The Pentium 4 - RIP?
Nick Maclaren
Does anyone else think that Intel's Banias CPU could well spell the
end of the line for the Pentium 4 - not in 2003, obviously, but in
the long term?
Server vendors are generally rather keener on the Pentium III than
the Pentium 4, and the current trendy type of system is the blade
server. My prediction is that we shall see OEMs start to produce
blade-based desktops (i.e. the CPU board of a desktop will be just
a server blade), and that purpose-designed desktops could well be
quite rare in a couple of years (being replaced by laptops and
blade-based systems).
If this is so, then the Prescott could well be Intel's last major
success with a primarily desktop chip, and any future Pentium 4
or successor would be squeezed out of the desktop market by the
"lower" end in the same way that the RISC chips were by the x86
range. My guess is that is what will happen, but I can't guess
whether Intel will attempt to produce a successor design (and I
don't just mean a new chip).
This then raises the interesting question of where Hyperthreading
is going, too. It could be put into the Banias design, or it
could be quietly dropped in favour of CMP.
Regards,
Nick Maclaren,
University of Cambridge Computing Service,
New Museums Site, Pembroke Street, Cambridge CB2 3QH, England.
Email: nm...@cam.ac.uk
Tel.: +44 1223 334761 Fax: +44 1223 334679
Terje Mathisen
> Does anyone else think that Intel's Banias CPU could well spell the
> end of the line for the Pentium 4 - not in 2003, obviously, but in
> the long term?
>
> Server vendors are generally rather keener on the Pentium III than
> the Pentium 4, and the current trendy type of system is the blade
> server. My prediction is that we shall see OEMs start to produce
> blade-based desktops (i.e. the CPU board of a desktop will be just
> a server blade), and that purpose-designed desktops could well be
> quite rare in a couple of years (being replaced by laptops and
> blade-based systems).
>
> If this is so, then the Prescott could well be Intel's last major
> success with a primarily desktop chip, and any future Pentium 4
> or successor would be squeezed out of the desktop market by the
> "lower" end in the same way that the RISC chips were by the x86
> range. My guess is that is what will happen, but I can't guess
> whether Intel will attempt to produce a successor design (and I
> don't just mean a new chip).
>
> This then raises the interesting question of where Hyperthreading
> is going, too. It could be put into the Banias design, or it
> could be quietly dropped in favour of CMP.
Nick, you disappoint me.
Given that profiling have shown Intel that 30-50 stage pipelines can be
a win, P4-style computing won't go away.
Blade servers are, imho, a solution looking for a problem.
What are they best at? MIPS/watt with limited RAM/cpu?
Terje
--
- <Terje.M...@hda.hydro.com>
"almost all programming can be viewed as an exercise in caching"
Tarjei T. Jensen
> Server vendors are generally rather keener on the Pentium III than
> the Pentium 4, and the current trendy type of system is the blade
> server. My prediction is that we shall see OEMs start to produce
> blade-based desktops (i.e. the CPU board of a desktop will be just
> a server blade), and that purpose-designed desktops could well be
> quite rare in a couple of years (being replaced by laptops and
> blade-based systems).
That may have started. I recently bought an epia m mainboard from VIA and a
suitable enclosure. The main board has a 933MHz x86 compatible C3 CPU. The
CPU does not have any of the SSE instructions. That has been rectified with
a new gigaHz CPU. With a CPU with SSE instructions, an epia based PC should
be useable for office work.
The power consumption of my PC is not much to write home about. It uses an
external powersupply which supplies 12V. The all the other voltages are
generated inside the case (separate circuit board). Some portables uses the
same external powersupply.
What I miss is a small form factor PC with ECC RAM. Machines like the
Sumicom s300 (http://www.kingyoung.com.tw/s300.htm) would be ideal for some
chores if it had ECC RAM. After reading in this newsgroup about defective
memory, I've become sceptical about using ordinary PCs as servers. I suppose
a server class CPU with internal ECC would be nice to have, but then the box
would perhaps not be cheap and cheerful anymore.
greetings,
Nick Maclaren
In article <b59pj0$fsj$1...@vkhdsu24.hda.hydro.com>,
Terje Mathisen <terje.m...@hda.hydro.com> writes:
|>
|> Nick, you disappoint me.
|>
|> Given that profiling have shown Intel that 30-50 stage pipelines can be
|> a win, P4-style computing won't go away.
When you don't like the answer, the solution is to change the
question! I am predicting that is what will happen - but, as
always when predicting the future, I could well be wrong.
|> Blade servers are, imho, a solution looking for a problem.
|>
|> What are they best at? MIPS/watt with limited RAM/cpu?
It depends on their interconnect. I am watching closely for any
blade servers that come out with an interesting interconnect.
This may never happen, of course.
Nick Maclaren
In article <3e786a0b$1...@news.wineasy.se>,
"Tarjei T. Jensen" <tarjei...@akerkvaerner.com> writes:
|> Nick Maclaren wrote:
|> > Server vendors are generally rather keener on the Pentium III than
|> > the Pentium 4, and the current trendy type of system is the blade
|> > server. My prediction is that we shall see OEMs start to produce
|> > blade-based desktops (i.e. the CPU board of a desktop will be just
|> > a server blade), and that purpose-designed desktops could well be
|> > quite rare in a couple of years (being replaced by laptops and
|> > blade-based systems).
|>
|> That may have started. I recently bought an epia m mainboard from VIA and a
|> suitable enclosure. The main board has a 933MHz x86 compatible C3 CPU. The
|> CPU does not have any of the SSE instructions. That has been rectified with
|> a new gigaHz CPU. With a CPU with SSE instructions, an epia based PC should
|> be useable for office work.
Interesting. We shall have to watch.
|> What I miss is a small form factor PC with ECC RAM. Machines like the
|> Sumicom s300 (http://www.kingyoung.com.tw/s300.htm) would be ideal for some
|> chores if it had ECC RAM. After reading in this newsgroup about defective
|> memory, I've become sceptical about using ordinary PCs as servers. I suppose
|> a server class CPU with internal ECC would be nice to have, but then the box
|> would perhaps not be cheap and cheerful anymore.
Yes, indeed. For anything that matters at all, ECC is essential.
Without it, HPC becomes a nightmare, because you have to run
every test several times :-(
Ketil Malde
> Given that profiling have shown Intel that 30-50 stage pipelines can
> be a win, P4-style computing won't go away.
Well, if you believe the Forrest curve, it may go "away" to a few
niches?
> Blade servers are, imho, a solution looking for a problem.
> What are they best at? MIPS/watt with limited RAM/cpu?
Surely MIPS/cubic metre? I thought the MIPS/Watt advantage was mainly
incidental.
-kzm
--
If I haven't seen further, it is by standing in the footprints of giants
Jakob Engblom
> Does anyone else think that Intel's Banias CPU could well spell the
> end of the line for the Pentium 4 - not in 2003, obviously, but in
> the long term?
Well, that would require the Banias to beat the desktop P4 on absolute
performance terms. And as long as the clock frequency on the P4 and
prescott ramps as predicted, that is not likely to happen. Banias
won't clock as highly due to its design it seems. 1.6 Ghz seems to be
a pretty tight max frequency for the current Banias, according to
reports, and Banias will likely require a transition to 90 nm design
rules to clock higher. So it cannot match a 3 Ghz hyperthreaded P4
currently.
Servers is certainly a market, and Intel could be big enough to have
two lines of processors: a max performance P4 line for desktop,
workstation, and gaming, and a lower-power, lower-performance laptop
and server line from Banias. Plus the Itanium, of course.
/jakob
Dennis M O'Connor
> Does anyone else think that Intel's Banias CPU could well spell the
> end of the line for the Pentium 4 - not in 2003, obviously, but in
> the long term?
No one with a real clue thinks that.
--
Dennis M. O'Connor dm...@primenet.com
Not speaking for Intel.
Bernd Paysan
> "Nick Maclaren" <nm...@cus.cam.ac.uk> wrote ...
>> Does anyone else think that Intel's Banias CPU could well spell the
>> end of the line for the Pentium 4 - not in 2003, obviously, but in
>> the long term?
>
> No one with a real clue thinks that.
Since Dennis (as usually) can't tell us Intel secrets, here's why:
The first Pentium 4 core (Willamette) had a significant die diet before it
shipped. This does have an impact on performance. I don't know how
significant it is on average code, but I won't be surprised if Prescott
cores would deliver a significantly higher IPC than previous cores.
On the other side: laptops and blades are really growing markets, unlike
desktops and server. A product like Banias, specifically targeted towards
laptops and blades is not a bad idea.
--
Bernd Paysan
"If you want it done right, you have to do it yourself"
http://www.jwdt.com/~paysan/
Allan Sandfeld Jensen
> Nick Maclaren wrote:
>> Server vendors are generally rather keener on the Pentium III than
>> the Pentium 4, and the current trendy type of system is the blade
>> server. My prediction is that we shall see OEMs start to produce
>> blade-based desktops (i.e. the CPU board of a desktop will be just
>> a server blade), and that purpose-designed desktops could well be
>> quite rare in a couple of years (being replaced by laptops and
>> blade-based systems).
>
> That may have started. I recently bought an epia m mainboard from VIA and
> a suitable enclosure. The main board has a 933MHz x86 compatible C3 CPU.
> The CPU does not have any of the SSE instructions. That has been rectified
> with a new gigaHz CPU. With a CPU with SSE instructions, an epia based PC
> should be useable for office work.
>
Sorry, what difference would SSE make? SSE is mostly useless, and for office
applications it is completely useless. The only applications to benefit
from it is multimedia and heavy graphics. The VIA Cyrix chips were designed
for specifically for typical office workload, and with that focus
implementing SSE made no sense. (hint: if your "CPU" is too slow for your
office applications try adding more RAM)
> The power consumption of my PC is not much to write home about. It uses an
> external powersupply which supplies 12V. The all the other voltages are
> generated inside the case (separate circuit board). Some portables uses
> the same external powersupply.
>
> What I miss is a small form factor PC with ECC RAM. Machines like the
> Sumicom s300 (http://www.kingyoung.com.tw/s300.htm) would be ideal for
> some chores if it had ECC RAM. After reading in this newsgroup about
> defective memory, I've become sceptical about using ordinary PCs as
> servers. I suppose a server class CPU with internal ECC would be nice to
> have, but then the box would perhaps not be cheap and cheerful anymore.
>
Checkout the Shuttle SV24, it sounds like what your are looking for.
greetings
`Allan
Michael S
> end of the line for the Pentium 4 - not in 2003, obviously, but in
> the long term?
>
No. Desktops are driven by gaming. Games are hungry for CPU cycles.
Pentium 4 delivers.
> Server vendors are generally rather keener on the Pentium III than
> the Pentium 4, and the current trendy type of system is the blade
> server. My prediction is that we shall see OEMs start to produce
> blade-based desktops (i.e. the CPU board of a desktop will be just
> a server blade), and that purpose-designed desktops could well be
> quite rare in a couple of years (being replaced by laptops and
> blade-based systems).
>
No. Economics of scale work in opposite direction.
Desktop->Workstation->Server.
> If this is so, then the Prescott could well be Intel's last major
> success with a primarily desktop chip, and any future Pentium 4
> or successor would be squeezed out of the desktop market by the
> "lower" end in the same way that the RISC chips were by the x86
> range. My guess is that is what will happen, but I can't guess
> whether Intel will attempt to produce a successor design (and I
> don't just mean a new chip).
>
It's possible, but the danger would come from the other low ends -
gaming consoles, set-top boxes and PDAs. I.e. from ARM and MIPS
cousins not from the x86 brother.
> This then raises the interesting question of where Hyperthreading
> is going, too. It could be put into the Banias design,
No. In the low power design it makes more sense to shut down unused
units then to try to reuse it.
> or it could be quietly dropped in favour of CMP.
>
No. CMP doesn't deliver in terms of SingleThreadedMIPS/$.
Hyperthreading does.
Greg Pfister
> In article <b59pj0$fsj$1...@vkhdsu24.hda.hydro.com>,
> Terje Mathisen <terje.m...@hda.hydro.com> writes:
> |> Blade servers are, imho, a solution looking for a problem.
> |>
> |> What are they best at? MIPS/watt with limited RAM/cpu?
>
> It depends on their interconnect. I am watching closely for any
> blade servers that come out with an interesting interconnect.
> This may never happen, of course.
It may not. On the other hand, it may. On the gripping hand, it may
already have happened: There are people who consider 100Mb Ethernet
quite adequately interesting, and are unwilling to pay for more than
that, since they don't need it. There are lots of web server farms out
there, and three-tier (web, app, DB) systems, and a fair number of
technical computing clusters that are quite happy with inexpensive
Ethernet as the interconnect. I'm not saying they're all deliriously
happy with current blade capabilities, but enough are happy to give
this form factor a hook into the market and pay for developing the
next generation.
One problem with producing systems like this with "interesting"
interconnects is that it's difficult to justify in terms of return on
development cost -- because the data isn't there to let you estimate
how many you might sell.
IDC, the primary purveyor of such data, has a single category for
technical computing, with no subdivisions. It doesn't distinguish the
sections of the market (and possible revenue from them) that need
faster interconnects from those that do. So there's massive
uncertainty here.
Don't assume that IDC is 100%, or maybe even 1%, to blame for that.
They report the data that people give them, and ask for (feedback
loop), and nobody, or at least not enough people, have asked for /
given the appropriate data.
--
Greg Pfister
Nick Maclaren
Bernd Paysan <bernd....@gmx.de> wrote:
>
>Since Dennis (as usually) can't tell us Intel secrets, here's why:
>
>The first Pentium 4 core (Willamette) had a significant die diet before it
>shipped. This does have an impact on performance. I don't know how
>significant it is on average code, but I won't be surprised if Prescott
>cores would deliver a significantly higher IPC than previous cores.
I was assuming that, but I believe that you and Jakob Engblom are
making a mistake in thinking that peak performance will be the main
issue for ever and a day. I think (but am not certain) that its
bubble is about to burst.
>On the other side: laptops and blades are really growing markets, unlike
>desktops and server. A product like Banias, specifically targeted towards
>laptops and blades is not a bad idea.
That is precisely the point. Look at where the Pentium 4 is used:
desktops and small servers, primarily.
It isn't implausible that its desktop market will shrink by 2-3 times
within a year or two, pressured by blades and laptops alone. And that
ignores the possibility of Athlon64, the POWER4 970 or any other new
competition.
Similarly, within that timescale, Intel will HAVE to get the Itanium
established in the small to medium server market or lose any chance
of the economies of scale. And there are serious competitors in that
market, too, to mention just the Opteron.
So the combination could easily see the Pentium 4 drop from being the
mass market leader to just another widespread chip (by Intel standards,
that is), with all appearances it it becoming a niche chip before long.
If that happens, I will place good money on Intel cancelling future
design projects.
Nick Maclaren
Greg Pfister <GPpf...@us.ibm.com.remove.GP.and.this.to.reply> wrote:
>
>It may not. On the other hand, it may. On the gripping hand, it may
>already have happened: There are people who consider 100Mb Ethernet
>quite adequately interesting, and are unwilling to pay for more than
Or even a 110 baud serial connexion :-) A few years ago, I got flamed
for saying that you could use postcards for communication, on a
sufficiently embarrassingly parallel task. Well, you can, and some
people did ....
Neither search problems nor Monte-Carlo methods are rare, and they
have essentially no communication requirements between initialisation
and termination.
>One problem with producing systems like this with "interesting"
>interconnects is that it's difficult to justify in terms of return on
>development cost -- because the data isn't there to let you estimate
>how many you might sell.
Yes, indeed. And, partly because there aren't enough cheap clusters
with high-performance interconnects, it isn't clear how many programs
that are currently serial or SMP-only could be converted to use them.
There's a hole in my bucket, dear Liza, dear Liza ....
Eric
> nm...@cus.cam.ac.uk (Nick Maclaren) wrote in message
> news:<b59lo4$sao$1...@pegasus.csx.cam.ac.uk>...
>> Does anyone else think that Intel's Banias CPU could well spell the
>> end of the line for the Pentium 4 - not in 2003, obviously, but in
>> the long term?
>>
>
> No. Desktops are driven by gaming. Games are hungry for CPU cycles.
> Pentium 4 delivers.
I thought the big problem was that desktops are not driven by anything these
days. A 3 year old computer, maybe with some added RAM, is fine for
everything I throw at it, including many games.
>> Server vendors are generally rather keener on the Pentium III than
>> the Pentium 4, and the current trendy type of system is the blade
>> server. My prediction is that we shall see OEMs start to produce
>> blade-based desktops (i.e. the CPU board of a desktop will be just
>> a server blade), and that purpose-designed desktops could well be
>> quite rare in a couple of years (being replaced by laptops and
>> blade-based systems).
>>
>
> No. Economics of scale work in opposite direction.
> Desktop->Workstation->Server.
That is only true after the "Desktop" gains some "Server"ish attributes.
And the tweaks necessary for a blade-type system (small form factor, good
thermal design, etc.) make sense for desktop hardware as well.
The last 5 systems I've bought have all been small form factor designs,
Shuttle "cube" systems or fanless VIA designs. I still have big, powerful
(expensive) workstation systems, but most have been kicked out and
relegated to server rooms where noise doesn't matter. Small, quiet boxes
make decent graphic terminals too.
>> If this is so, then the Prescott could well be Intel's last major
>> success with a primarily desktop chip, and any future Pentium 4
>> or successor would be squeezed out of the desktop market by the
>> "lower" end in the same way that the RISC chips were by the x86
>> range. My guess is that is what will happen, but I can't guess
>> whether Intel will attempt to produce a successor design (and I
>> don't just mean a new chip).
I don't see how Intel's marketing could deal with such a processor. Saying
"you don't need the best and fastest" is not a compelling reason to
upgrade.
>> This then raises the interesting question of where Hyperthreading
>> is going, too. It could be put into the Banias design,
>
> No. In the low power design it makes more sense to shut down unused
> units then to try to reuse it.
Not if there is going to be One True CPU. In heat/power-conscious
environments (laptop on batteries, blade server, anything where the
temperature is too high), disable threads 2..N and shut down idle units.
Probably would use a little more power than a non-HT CPU for laptops, and
be a little less efficient than a pure-HT system for desktops/servers. But
the observation Nick made (perhaps I am misreading) was this:
Maybe it is ok to trade 5-10% performance in exchange for something that
works everywhere.
I think he's right.
-Eric
Bill Todd
"Tarjei T. Jensen" <tarjei...@akerkvaerner.com> wrote in message
news:3e786a0b$1...@news.wineasy.se...
...
After reading in this newsgroup about defective
> memory, I've become sceptical about using ordinary PCs as servers. I
suppose
> a server class CPU with internal ECC would be nice to have, but then the
box
> would perhaps not be cheap and cheerful anymore.
ECC PC memory itself is only slightly more expensive than non-ECC PC memory.
Motherboards supporting it are available at prices only slightly higher than
vanilla PC MBs. And it's not clear that there's any substantive difference
in reliability between a P4 and a Xeon, particularly if you underclock the
P4 a tad.
I suspect that you could build an industrial-strength server out of the same
parts (CPU perhaps excluded) that Dell uses and save a considerable amount
of cash, or save even more by investigating what parts might be available
from other sources with no substantive loss in quality (though of course
your time would cost something). Of course, careful assembly and thorough
burn-in and component-compatibility testing are required in any event.
- bill
Andrew Reilly
> Blade servers are, imho, a solution looking for a problem.
>
> What are they best at? MIPS/watt with limited RAM/cpu?
MIPS/m^3 might be (seen to be) the main technical advantage, but I think
that in many situations they are a response to a /software/ management
and reliability problem. You get partitioning of your applications that
today's operating systems can't guarantee, you get to do maintainence
without taking your whole operation down, you get a measure of fault
tollerance and inexpensive incremental growth.
Like all software problems, it's easy to argue that it can be fixed in
software.
--
Andrew
Peter da Silva
> No. Desktops are driven by gaming. Games are hungry for CPU cycles.
> Pentium 4 delivers.
Multiprocessors deliver CPU cycles better, and it's getting closer
to the time when multiprocessing is the most efective way of
making faster systems... if it wasn't for the Compaqtion, Alpha
would possibly have beeen following that thread to higher
perfomance by now. It's just a matter of coding games for them,
and then MIPS/watt becomes more interesting.
http://www.theinquirer.net/?article=7078
--
`-_-' Ar rug tú barróg ar do mhactíre inniu?
'U` Peter da Silva
Peter da Silva
> My prediction is that we shall see OEMs start to produce
> blade-based desktops (i.e. the CPU board of a desktop will be just
> a server blade), and that purpose-designed desktops could well be
> quite rare in a couple of years (being replaced by laptops and
> blade-based systems).
One hopes that they're not quite as tightly integrated as blades.
I would love a backplane based desktop rather than a motherboard
based one, but having everything integrated with the processor
card drops you right back in the same forklift-upgrade-only world.
Jan Mikkelsen
>Blade servers are, imho, a solution looking for a problem.
>
>What are they best at? MIPS/watt with limited RAM/cpu?
MIPS/RU and MIPS/$
Commercial application server farms is the primary application I see,
where 1RU dual P3 systems are currently used. A blade lets you have
more MIPS per RU and you don't pay for stuff you don't need (SCSI
hotswap drives, PCI slots). The entire blade is the hotswap unit.
Regards,
Jan Mikkelsen
ja...@transactionware.com
Robert Myers
wrote:
>
>http://www.theinquirer.net/?article=7078
Ummm, a *teraflop*? In a game console? Is this for real?
RM
Stephen Fuld
"Greg Pfister" <GPpf...@us.ibm.com.remove.GP.and.this.to.reply> wrote in
message news:3e78a...@news1.prserv.net...
snip
> One problem with producing systems like this with "interesting"
> interconnects is that it's difficult to justify in terms of return on
> development cost -- because the data isn't there to let you estimate
> how many you might sell.
>
> IDC, the primary purveyor of such data, has a single category for
> technical computing, with no subdivisions. It doesn't distinguish the
> sections of the market (and possible revenue from them) that need
> faster interconnects from those that do. So there's massive
> uncertainty here.
>
> Don't assume that IDC is 100%, or maybe even 1%, to blame for that.
> They report the data that people give them, and ask for (feedback
> loop), and nobody, or at least not enough people, have asked for /
> given the appropriate data.
One must suppose that the powers that be at IBM are sure enough that there
isn't a market that they won't spend the modest amount of money to
commission a special study (by IDC or any of several other competent
companies) to get the information. I would guess you could get that
information for well under $100K in market research dollars.
--
- Stephen Fuld
e-mail address disguised to prevent spam
Kjetil Torgrim Homme
it's single precision. the PS2 has 6 Gflops by the same measure,
which put the Pentium to shame when it was released. it is not
unrealistic for games to actually achieve close to those 6 Gflops
(physics and 3D transformations), but if used for general computing it
won't come anything near that.
--
Kjetil T. | read and make up your own mind
| http://www.cactus48.com/truth.html
McCalpin
Greg Pfister <GPpf...@us.ibm.com.remove.GP.and.this.to.reply> wrote:
>
>technical computing, with no subdivisions.
This is not exactly true....
IDC's widely distributed reports break down "High Performance
Computing" revenue by price band (four tiers) and by vendor.
For suitably large piles of $$$, it is also possible to obtain
the IDC "pivot" spreadsheets which include quarterly estimates of
sales quantity and revenue for each model from each vendor.
These allows quantitative review of the breakdown between cluster
sales and single SMP sales in each price band, and allow us to
monitor the vector vs NUMA vs MPP vs "other" split in the top end
of the HPC market space.
IDC occasionally does more detailed reports, such as a detailed
breakdown of the HPC market by application segment a few years
ago. Each application segment was reviewed by price band,
vendor revenue, growth rates, and other factors.
IDC also publishes reports on the cluster market, but these are
brutally difficult to interpret because of uncertainties in the
base data. (Clusters can be acquired in many ways, and the only
ones that are easy to track are the clusters with unique model
numbers from the large vendors.)
Finally, IDC does detailed specialty reports for customers.
I don't think that it is giving away any big secrets to say that
IBM has commissioned some of these over the years.
Interpreting IDC reports is something of an art, since IDC's data
comes from vendors, and there is a certain "creativity" in what
vendors tell IDC (just like there is a certain "creativity" in
submitting systems to the "TOP500 List of Supercomputers").
IDC also has different rules for reporting income from hardware,
software, services, and consulting than the vendors have
internally, so sometimes it is a lot of work to bridge between
company internal numbers and IDC numbers. Despite these caveats,
there is no doubt that the IDC market data for HPC is a unique
and extremely valuable resource.
--
John D. McCalpin, Ph.D. mcca...@austin.ibm.com
Senior Technical Staff Member IBM POWER Microprocessor Development
"I am willing to make mistakes as long as
someone else is willing to learn from them."
Mitch Alsup
> Nick Maclaren wrote:
> > Does anyone else think that Intel's Banias CPU could well spell the
> > end of the line for the Pentium 4 - not in 2003, obviously, but in
> > the long term?
Not a chance--although one should be left with hope.....
> >
> > Server vendors are generally rather keener on the Pentium III than
> > the Pentium 4, and the current trendy type of system is the blade
> > server. My prediction is that we shall see OEMs start to produce
> > blade-based desktops (i.e. the CPU board of a desktop will be just
> > a server blade), and that purpose-designed desktops could well be
> > quite rare in a couple of years (being replaced by laptops and
> > blade-based systems).
One reason for the PIII servers is the lower cost of the memory systems,
another is the lower latency of PIII bus accesses when compared with P4.
> >
> > If this is so, then the Prescott could well be Intel's last major
> > success with a primarily desktop chip, and any future Pentium 4
> > or successor would be squeezed out of the desktop market by the
> > "lower" end in the same way that the RISC chips were by the x86
> > range. My guess is that is what will happen, but I can't guess
> > whether Intel will attempt to produce a successor design (and I
> > don't just mean a new chip).
> >
> > This then raises the interesting question of where Hyperthreading
> > is going, too. It could be put into the Banias design, or it
> > could be quietly dropped in favour of CMP.
>
> Nick, you disappoint me.
>
> Given that profiling have shown Intel that 30-50 stage pipelines can be
> a win, P4-style computing won't go away.
>
> Blade servers are, imho, a solution looking for a problem.
>
> What are they best at? MIPS/watt with limited RAM/cpu?
>
> Terje
Given that major high volume x86 machines are being designed with a
power budget that looks like: (with all the appropriate caveats)
clock: 40%
leakage: 20%
flops: 20%
gates: 20%
A) increasing the depth of the pipeline can indeed gain small
performance advantages.
B) However, Doubling the flop count to use the gates more often
often increases power at a much faster rate than performance.
Of course at this point I am comparing a P4-like pipeline with a
pipeline with tiwce as many stages. Due to the overhead of clock
jitter, flop delay, and clock skew; this pipe should run on-the-
order-of only 50% faster than a P4 from 1/2 as many logic gates
per cycle. I.e. 80% more power 50% faster frequency, and a lot
of trouble holding on to the IPC.
C) the depth to which a CPU can be pipelined is a function of
the branch prediction and branch density. When a longer pipe
machine is only a little bit faster on todays applications, it
can end up slightly slower on tomorrows application.
Mitch Alsup
Patrick Schaaf
>Terje Mathisen <terje.m...@hda.hydro.com> wrote:
>>Blade servers are, imho, a solution looking for a problem.
>>
>>What are they best at? MIPS/watt with limited RAM/cpu?
>MIPS/RU and MIPS/$
Can you give a concrete example where a currently available blade
is compared to an equal number of same-RAM / same-Processors
1U systems, and comes out with a better MIPS/$? I have not seen
one, yet.
Personal reasons why no blade I've seen appeals to me:
1) System management console (extra ethernet) not per individual blade.
You can usually switch back and forth, but no concurrent operation
on different blade appears possible, and the security model is lacking.
2) Some internal Ethernet switch that I have to learn to manage.
We run a tight, integrated cloud of ethernet switches, centrally
managed, with very differentiated per-port VLAN settings. We use
Cisco Switches (Cat650x in the core, 2980 for aggregation).
Each blade server comes with _some_ switch, all have their
limitations, and would need careful integration, if possible
at all.
Both criteria are central for us: we build systems into racks, and
LATER parcel them out to one or the other administrator, after setting
them up (anew, if it's a project switch), and after putting them into
the desired VLAN. Those other administrators work remotely, per default,
and need individual access to a management console.
We grew our operation, keeping racks and cabling in a very pure state,
by having the flexibility to NOT force a set of commonly administered
systems onto one switch / one rack / one shared management card.
We don't want to part with that flexibility.
Note that this is NOT the classical housing data center. Almost all machines
provide services for our own company, and are owned by us. They are just
developed, administered, and deployed by various small teams of people,
each team responsible for the (web/db) applications on one or two
groups of systems.
best regards
Patrick
Nick Maclaren
In article <b5ar6v$1fig$2...@jeeves.eng.abbnm.com>,
Peter da Silva <pe...@abbnm.com> writes:
|> Nick Maclaren wrote:
|> > My prediction is that we shall see OEMs start to produce
|> > blade-based desktops (i.e. the CPU board of a desktop will be just
|> > a server blade), and that purpose-designed desktops could well be
|> > quite rare in a couple of years (being replaced by laptops and
|> > blade-based systems).
|>
|> One hopes that they're not quite as tightly integrated as blades.
|> I would love a backplane based desktop rather than a motherboard
|> based one, but having everything integrated with the processor
|> card drops you right back in the same forklift-upgrade-only world.
Given that a decreasing proportion of the market is to people (or
even retailers) who can or will replace anything smaller than a
major board, I am afraid that you are out of luck. Note that I am
not saying that this is a desirable change - merely that it is the
situation.
Nick Maclaren
In article <PF8ea.18420$S%3.10...@bgtnsc04-news.ops.worldnet.att.net>,
"Stephen Fuld" <s.f...@PleaseRemove.att.net> writes:
|> "Greg Pfister" <GPpf...@us.ibm.com.remove.GP.and.this.to.reply> wrote in
|> message news:3e78a...@news1.prserv.net...
|>
|> > interconnects is that it's difficult to justify in terms of return on
|> > development cost -- because the data isn't there to let you estimate
|> > how many you might sell.
|>
|> isn't a market that they won't spend the modest amount of money to
|> commission a special study (by IDC or any of several other competent
|> companies) to get the information. I would guess you could get that
|> information for well under $100K in market research dollars.
Regrettably, such information is worth very much less than you
pay for it! IDC have at best a poor feel for innovative markets,
and it isn't clear who has a better one. IBM could consult someone
like me, whose predictions would be less based on yesterday's model,
but would run the corresponding risk of being completely off-beam.
There was a classic example of using IDC-like surveys just before
the actual release of the PowerPC (the system, I mean, not the chip).
IBM consulted someone who consulted management, who said that it
was absolutely essential to be able to reuse their ISA disks from
the desktops they were replacing. IBM then changed the PowerPC from
a SCSI-only system to an ISA one, with SCSI as an expensive optional
extra.
I told the head of marketing in the UK that IBM had got this badly
wrong, and he told me that the above was the feedback that IBM had
got from customers. Well, history relates that the original PowerPC
designers and I were 100% right and the "customer survey" was 99%
wrong.
Tarjei T. Jensen
> Sorry, what difference would SSE make? SSE is mostly useless, and for
office
> applications it is completely useless. The only applications to benefit
> from it is multimedia and heavy graphics. The VIA Cyrix chips were
designed
> for specifically for typical office workload, and with that focus
> implementing SSE made no sense. (hint: if your "CPU" is too slow for your
> office applications try adding more RAM)
I am not convinced that this theory holds water. I suspect that there are
small corners where the SSE instruction are very useful, but it won't show
up in the big picture.
BTW I consider Windows Media Player an office application.
Anyway my system has 512MB RAM.
> Checkout the Shuttle SV24, it sounds like what your are looking for.
I intend to.
greetings,
Nick Maclaren
In article <b5btaa$o58$1...@pegasus.csx.cam.ac.uk>,
nm...@cus.cam.ac.uk (Nick Maclaren) writes:
|>
|> There was a classic example of using IDC-like surveys just before
|> the actual release of the PowerPC (the system, I mean, not the chip).
|> IBM consulted someone who consulted management, who said that it
|> was absolutely essential to be able to reuse their ISA disks from
|> the desktops they were replacing. IBM then changed the PowerPC from
|> a SCSI-only system to an ISA one, with SCSI as an expensive optional
|> extra.
I mean IDE, not ISA, of course. TLA overload strikes again!
Michael S
> Michael S wrote:
>
> > nm...@cus.cam.ac.uk (Nick Maclaren) wrote in message
> > news:<b59lo4$sao$1...@pegasus.csx.cam.ac.uk>...
> >> is going, too. It could be put into the Banias design,
> >
> > No. In the low power design it makes more sense to shut down unused
> > units then to try to reuse it.
>
> Not if there is going to be One True CPU. In heat/power-conscious
> environments (laptop on batteries, blade server, anything where the
> temperature is too high), disable threads 2..N and shut down idle units.
> Probably would use a little more power than a non-HT CPU for laptops, and
> be a little less efficient than a pure-HT system for desktops/servers. But
> the observation Nick made (perhaps I am misreading) was this:
>
> Maybe it is ok to trade 5-10% performance in exchange for something that
> works everywhere.
>
> I think he's right.
>
> -Eric
If we were talking about other company - may be. But we are talking
Intel. Intel, Motorola and TI never were obsessed with the "one size
fits all" idea. They always had many architectures combined with many
implementation for each architecture. Think Intel, 1990 - one 8-bitter
(8051), two 16-bitters (186, 196), three 32-bitters (386, 960, 860).
Each one (with exciption of 186 and 860) has multiple implementations.
Actually, I was surprised when Intel started to drop 960 line
recently. It was very Intelish to keep both 960 and StrongArm/Xscale
competing on the same market space.
Nick Maclaren
In article <f881b862.03032...@posting.google.com>,
already...@yahoo.com (Michael S) writes:
|>
|> If we were talking about other company - may be. But we are talking
|> Intel. Intel, Motorola and TI never were obsessed with the "one size
|> fits all" idea. They always had many architectures combined with many
|> implementation for each architecture. Think Intel, 1990 - one 8-bitter
|> (8051), two 16-bitters (186, 196), three 32-bitters (386, 960, 860).
|> Each one (with exciption of 186 and 860) has multiple implementations.
|> Actually, I was surprised when Intel started to drop 960 line
|> recently. It was very Intelish to keep both 960 and StrongArm/Xscale
|> competing on the same market space.
That's true, but I was not referring to Intel's decisions. Intel
may like selling lots of lines, but most OEMs are less keen, and
peripheral manufacturers and ISVs positively loathe them. Especially
given the quite major differences between the Pentium M and the HT
Pentium 4, I am not expecting to see a huge number of peripherals
and applications tuned for both - and some won't even be supported
for both.
Also, note that Intel are going to have to cut prices and market
the Pentium M aggressively if they want to maintain dominance in
the blade market against AMD, Transmeta, Sun and even IBM. Which
is going to reduce the price advantages of the Pentium 4.
Lastly, Intel is in the position of a pack leader that has missed
its kill. Intel has been forced to back down by OEMs several times
in the past couple of years - unthinkable at the peak of their
dominance - and so is vulnerable to pressure from them.
Thomas Womack
Nick Maclaren <nm...@cus.cam.ac.uk> wrote:
> That's true, but I was not referring to Intel's decisions. Intel
> may like selling lots of lines, but most OEMs are less keen, and
> peripheral manufacturers and ISVs positively loathe them.
Why do peripheral manufacturers care? The difference between CPUs and
between chipsets is pretty much imperceptible at the IDE bus, the
USB socket or even the PCI slot.
>Also, note that Intel are going to have to cut prices and market
>the Pentium M aggressively if they want to maintain dominance in
>the blade market against AMD, Transmeta, Sun and even IBM.
I'm not sure that's true either. The "blade market", whose existence
I rather doubt, can't be price-sensitive given how much of a premium
is charged for the machines; and the Pentium M puts Intel in a
different league to Transmeta or Sun for performance in their blades,
and a different league to AMD and IBM for power consumption.
I'd be interested to see Pentium M SPEC benchmarks: I'm expecting
something comparable to a 2400MHz Xeon.
Tom
Nick Maclaren
In article <b5civq$2ta$1...@oyez.ccc.nottingham.ac.uk>, pmx...@merlot.uucp (Thomas Womack) writes:
|> In article <b5cia4$d1g$1...@pegasus.csx.cam.ac.uk>,
|> Nick Maclaren <nm...@cus.cam.ac.uk> wrote:
|>
|> > That's true, but I was not referring to Intel's decisions. Intel
|> > may like selling lots of lines, but most OEMs are less keen, and
|> > peripheral manufacturers and ISVs positively loathe them.
|>
|> Why do peripheral manufacturers care? The difference between CPUs and
|> between chipsets is pretty much imperceptible at the IDE bus, the
|> USB socket or even the PCI slot.
Think interrupt handling and drivers. Think validation and support.
It is extremely common for the higher-RAS systems to support only
those combinations that they have actually validated, and this
applies equally to peripherals and software.
|> >Also, note that Intel are going to have to cut prices and market
|> >the Pentium M aggressively if they want to maintain dominance in
|> >the blade market against AMD, Transmeta, Sun and even IBM.
|>
|> I'm not sure that's true either. The "blade market", whose existence
|> I rather doubt, can't be price-sensitive given how much of a premium
|> is charged for the machines; and the Pentium M puts Intel in a
|> different league to Transmeta or Sun for performance in their blades,
|> and a different league to AMD and IBM for power consumption.
When things are new and trendy, vendors charge premiums, but I can
assure you that the blade market both exists and will be the scene
for some serious competition. Just watch it :-)
Your last point is true, at present, but may not be so indefinitely,
and then there is the question of serious discounting. Is it better
to buy N blades of performance P or 3N blades of performance P/2?
|> I'd be interested to see Pentium M SPEC benchmarks: I'm expecting
|> something comparable to a 2400MHz Xeon.
Agreed, there, which is why I am interested in the CPU.
David Wang
> I'd be interested to see Pentium M SPEC benchmarks: I'm expecting
> something comparable to a 2400MHz Xeon.
Get someone to lend me a box for a month. I'll do a complete
run down, including SPEC and compared it to a 2.4 GHz P4.
Perhaps the editor of RWT has contacts to get loaners. :)
--
davewangATwamDOTumdDOTedu
Robert Harley
Terje Mathisen <terje.m...@hda.hydro.com> writes:
> Given that profiling have shown Intel that 30-50 stage pipelines can
> be a win, P4-style computing won't go away.
Then again,
http://www.intel.com/products/benchmarks/notebook/
suggests that a tweaked PIII @ 1.6 GHz easily beats a P4 @ 2.4 GHz,
and would at least match a hypothetical 2.75 GHz one (on BAPCo MobileMark).
R
Nick Maclaren
In article <rz77kau...@jaune.inria.fr>,
It's going to be an amusing year, isn't it? First work out what is
competing with what :-)
Terje Mathisen
So, what this test is saying is that Microsoft Office and similar apps
are so much more branchy than SPEC, that a shorter pipeline length might
be closer to optimum?
Not too surprising, but still interesting.
Terje
--
- <Terje.M...@hda.hydro.com>
"almost all programming can be viewed as an exercise in caching"
Greg Pfister
> In article <3e78a...@news1.prserv.net>,
> Greg Pfister <GPpf...@us.ibm.com.remove.GP.and.this.to.reply> wrote:
>
>>It may not. On the other hand, it may. On the gripping hand, it may
>>already have happened: There are people who consider 100Mb Ethernet
>>quite adequately interesting, and are unwilling to pay for more than
>>that, since they don't need it. ...
>
>
> Or even a 110 baud serial connexion :-) A few years ago, I got flamed
> for saying that you could use postcards for communication, on a
> sufficiently embarrassingly parallel task. Well, you can, and some
> people did ....
I usually refer to connection using wet string. :-) I don't get flamed
for it, though, even though I've used that for a number of years. Must
be the delivery. :-)
> Neither search problems nor Monte-Carlo methods are rare, and they
> have essentially no communication requirements between initialisation
> and termination.
Completely true. It's amusing to consider simulations running huge
numbers of regression test cases as search, but I guess it is. When I
put the "cure for Anthrax" UD background task on my laptop, I was
purely grossed out at how inefficient that was as a search.
--
Greg Pfister
Greg Pfister
> Michael S wrote:
>
>>No. Desktops are driven by gaming. Games are hungry for CPU cycles.
>>Pentium 4 delivers.
>
>
> Multiprocessors deliver CPU cycles better,
? Define "better."
> and it's getting closer
> to the time when multiprocessing is the most efective way of
> making faster systems...
There are always cases requiring more processing than any
"multiprocessor" box can provide. This depends, of course, on how you
define "multiprocessor." NUMA systems with very large nonuniformity
fuzz the difference.
> if it wasn't for the Compaqtion,
Good term! I'm going to use it. Is it original with you? (I like
attribution.)
--
Greg Pfister
Greg Pfister
Traditionally that's been the case. The tide may be turning, slowly.
But don't lay blame solely on IBM. Anybody else could have asked the
same questions.
And it's not just a question of the specific cost you mention. IDC
can't analyze & publish data it can't get. It gets the data by asking
companies -- and if they don't have their data broken out that way,
it's simply not available. There's significant cost in the data
gathering, too, much more than the amount mentioned when you consider
doing it for year after year.
--
Greg Pfister
Stephen Fuld
"Nick Maclaren" <nm...@cus.cam.ac.uk> wrote in message
news:b5btaa$o58$1...@pegasus.csx.cam.ac.uk...
>
> In article <PF8ea.18420$S%3.10...@bgtnsc04-news.ops.worldnet.att.net>,
> "Stephen Fuld" <s.f...@PleaseRemove.att.net> writes:
> |> "Greg Pfister" <GPpf...@us.ibm.com.remove.GP.and.this.to.reply> wrote
in
> |> message news:3e78a...@news1.prserv.net...
> |>
> |> > One problem with producing systems like this with "interesting"
> |> > interconnects is that it's difficult to justify in terms of return on
> |> > development cost -- because the data isn't there to let you estimate
> |> > how many you might sell.
> |>
> |> One must suppose that the powers that be at IBM are sure enough that
there
> |> isn't a market that they won't spend the modest amount of money to
> |> commission a special study (by IDC or any of several other competent
> |> companies) to get the information. I would guess you could get that
> |> information for well under $100K in market research dollars.
>
> Regrettably, such information is worth very much less than you
> pay for it! IDC have at best a poor feel for innovative markets,
> and it isn't clear who has a better one.
There are certainly perils from taking such information at face value,
especially for innovative products, but if you know that going in, you can
compensate to some extent for it.
snipped your "classic example"
When Chester Carlson came to IBM with the idea for a machine that would
allow one to make an arbitrary number of copies of a document typed on
ordinary paper, they talked to many of the secretaries within IBM and asked
them how often they made more than five copies of a typed document. They
answered essentially never because they could get five copies with carbon
paper and that was about the limit so they rarely did more. (If they knew
they wanted more, they would use a ditto or mimeograph stencil.) As a
result of this "market research" IBM told Carlson they weren't interested.
Carlson then took his idea to Halide who took the chance and Xerox was born.
Paul DeMone
Robert Harley wrote:
The Banias is an impressive design all right considering its perfomance/area
or performance/Watt. OTOH it is newer than P4 and very likely isn't as
scalable in absolute performance as P4 in the same technology.
Why does the fact that different commercial microarchitectures implemented
in similar process technology (more or less identical in this case) and design
constraints generate similar levels of performance seem so surprising?
Think back to the great rivalry between HP and DEC in the 1990s. How
generally close the performance race was between the PA-7x00 and Alpha
EV4, and later between the PA-8x00 and EV5x despite absolutely huge
differences in implementation approaches - microarchitecture and of
course clock rates.
If there was one approach to designing microprocessors that was absolutely
and clearly superior to all others for all applications everyone would be using
it.
Aside from the technological argument, consider the process of natural
selection. If Banias couldn't offer performance comparable or better than
Pentium 4M in a power constrained target platform then it wouldn't have
been developed and brought to market. OTOH, if the P4 microarchitecture
can't provide better performance for desktop form factor than a short pipe
design Intel could have succeeded the P6 core with a Banias style design
and saved itself the expense of building a good fraction of its 0.13 um fab
capacity.
--
Paul W. DeMone The 801 experiment SPARCed an ARMs race of EPIC
Kanata, Ontario proportions to put more PRECISION and POWER into
pde...@igs.net architectures with MIPSed results but ALPHA's well
that ends well.
Paul DeMone
Terje Mathisen wrote:
> > suggests that a tweaked PIII @ 1.6 GHz easily beats a P4 @ 2.4 GHz,
> > and would at least match a hypothetical 2.75 GHz one (on BAPCo MobileMark).
>
> So, what this test is saying is that Microsoft Office and similar apps
> are so much more branchy than SPEC, that a shorter pipeline length might
> be closer to optimum?
>
> Not too surprising, but still interesting.
Studies I have seen suggest PC productivity apps are less sensitive to
branch misprediction penalty than SPECint 2k and far less sensitive
than SPECint 95.
Greg Pfister
[snip]
> When Chester Carlson came to IBM with the idea for a machine that would
> allow one to make an arbitrary number of copies of a document typed on
> ordinary paper, they talked to many of the secretaries within IBM and asked
> them how often they made more than five copies of a typed document. They
> answered essentially never because they could get five copies with carbon
> paper and that was about the limit so they rarely did more. (If they knew
> they wanted more, they would use a ditto or mimeograph stencil.) As a
> result of this "market research" IBM told Carlson they weren't interested.
> Carlson then took his idea to Halide who took the chance and Xerox was born.
Classic story about the founding of Xerox. I don't know whether it's
true. I have also heard (Nth hand, rumour) that the guy who invented
the selectric ball was fired and re-hired some large number of times
before he got the thing produced. I've no idea whether that was
related to market research; could have been personal characteristics.
As others have pointed out, market reserach is a whole lot better at
predicting an existing market, where there's a historical perspective,
than at predicting new markets, where there's no easily quantifiable
data and attempts to get it can be polluted by, e.g., less than
persuasive descriptions to potential users -- or excessively
persuasive descriptions.
--
Greg Pfister
Greg Lindahl
Terje Mathisen <terje.m...@hda.hydro.com> wrote:
>So, what this test is saying is that Microsoft Office and similar apps
>are so much more branchy than SPEC, that a shorter pipeline length might
>be closer to optimum?
SPECcpu is composed of a bunch of different apps with different
characteristics. What does your statement actually mean?
greg
Terje Mathisen
Just what I said, Greg:
Since the performance ratio between PIII and P4 class cpus vary a lot
between SPEC on one side and BAPCO (i.e. mostly Office) on the other
side, the longer pipeline in the P4 must cost more for those apps.
Do you have any other explanation for the performance differences?
Nick Maclaren
Greg Pfister <GPpf...@us.ibm.com.remove.GP.and.this.to.reply> wrote:
>Nick Maclaren wrote:
>> In article <3e78a...@news1.prserv.net>,
>> Greg Pfister <GPpf...@us.ibm.com.remove.GP.and.this.to.reply> wrote:
>>
>>>It may not. On the other hand, it may. On the gripping hand, it may
>>>already have happened: There are people who consider 100Mb Ethernet
>>>quite adequately interesting, and are unwilling to pay for more than
>>>that, since they don't need it. ...
>>
>> Or even a 110 baud serial connexion :-) A few years ago, I got flamed
>> for saying that you could use postcards for communication, on a
>> sufficiently embarrassingly parallel task. Well, you can, and some
>> people did ....
>
>I usually refer to connection using wet string. :-) I don't get flamed
>for it, though, even though I've used that for a number of years. Must
>be the delivery. :-)
Yes, it was :-) The aspect of the delivery that wound up the flamers
was that I was using a reductio ad absurdum argument on their claims!
Intellectually justified, but I agree not tactful.
Anne & Lynn Wheeler
> There are certainly perils from taking such information at face
> value, especially for innovative products, but if you know that
> going in, you can compensate to some extent for it.
>
> snipped your "classic example"
we weere doing things with T1s in the early to mid '80s ... and had
some installed. The (official) communication division had product
that only supported multiple 56kbit links. they commisioned a study
that asked customers how many "parallel" 56kbit links that they were
running between two locations. The results showed no data points with
over five 56kbit links between (same) two locations. Based on the
results, they concluded that they wouldn't need a product supporting T1
for another 8-10 years.
Note 1: they didn't ask how many were running T1 between two
sites (using some other vendor's product); a trivial survey found
200.
Note 2: they didn't ask what the tariff was for T1; the avg. T1 tariff
tended to be somewhere around 5-7 individual 56kbit links. one of the
reasons that there were no data points with over five (parallel)
56kbit links, was that the tarrif for a T1 tended to be cheaper than
the tarrif for six (parallel) 56kbit links.
random refs:
http://www.garlic.com/~lynn/subtopic.html#hsdt
then there was the workstation review at PASC
http://www.garlic.com/~lynn/96.html#4a John Hartmann's Birthday Party
http://vm.marist.edu/~piper/party/jph-12.html#wheeler
--
Anne & Lynn Wheeler | http://www.garlic.com/~lynn/
Internet trivia 20th anv http://www.garlic.com/~lynn/rfcietff.htm
Anton Ertl
>Since the performance ratio between PIII and P4 class cpus vary a lot
>between SPEC on one side and BAPCO (i.e. mostly Office) on the other
>side, the longer pipeline in the P4 must cost more for those apps.
>
>Do you have any other explanation for the performance differences?
SPEC is provided as source, and BapCo provided as binaries; the SPEC
numbers are produced with Intel's C compiler, with Pentium 4 specific
optimizations, many of the BapCo binaries are produced with other
compilers, and not necessarily with optimzations for Pentium 4 (if the
compiler version that was used for the binary already knew about the
Pentium 4 at all).
- anton
--
M. Anton Ertl Some things have to be seen to be believed
an...@mips.complang.tuwien.ac.at Most things have to be believed to be seen
http://www.complang.tuwien.ac.at/anton/home.html
Del Cecchi
the market for even new models. Here is an example from personal experience: At
the time of the B models of AS400,long about 1990, we came up with an idea that
was heresy at the time. We were going to sort AS400 IMPI processors made out of
Bipolar Gate Arrays by picking out the fast parts. We were able to reduce the
processor cycle time from 60 ns to 45, as I recall. I told you it was a long
time ago.
The market forecasting folk predicted we would sell a couple of hundred over the
life of the program. Actually we sold over a thousand the first year.
For an equivilent story, see the 3033.
del cecchi
--
Del Cecchi
cec...@us.ibm.com
Personal Opinions Only
Jan Mikkelsen
Patrick Schaaf <mailer...@bof.de> wrote:
>ja...@godzilla.zeta.org.au (Jan Mikkelsen) writes:
>
>>Terje Mathisen <terje.m...@hda.hydro.com> wrote:
>>>Blade servers are, imho, a solution looking for a problem.
>>>
>>>What are they best at? MIPS/watt with limited RAM/cpu?
>
>>MIPS/RU and MIPS/$
>
>Can you give a concrete example where a currently available blade
>is compared to an equal number of same-RAM / same-Processors
>1U systems, and comes out with a better MIPS/$? I have not seen
>one, yet.
I went to the Australian IBM site (www.ibm.com.au) are did a comparison
between the x335 (dual P4-2400) and the HS20 blade and chassis. I'm
not advocating IBM; they just happened to be there.
(All prices are AUD, include 10% tax, retail off the web)
x335, 2 x Xeon-2400, 1 x 36 GB SCSI drive, 1.5GB RAM =
$9685 each.
Cost for a 84 CPUs (42U) =
42 * $9685 =
$406770
HS20, 2 x Xeon-2400, 1 x 40GB IDE drive, 1.5GB RAM =
$8985 each.
Chassis, holds 14 blades (7RU) =
$7919
Cost for a 84 CPUs (21RU) =
3 * $7919 + 3 * 14 * $8985 =
$401127
So, the MIPS/$ is essentially identical. However, the
blade approach was half the price in terms of rack units consumed.
Where floorspace is expensive, this can add up.
>Personal reasons why no blade I've seen appeals to me:
>
>1) System management console (extra ethernet) not per individual blade.
> You can usually switch back and forth, but no concurrent operation
> on different blade appears possible, and the security model is lacking.
I don't see this as a big issue, but it might be important for you.
Even 1RU servers typically have a switch where only one person has
"console" access at a given time.
My take on this is that once the operating system is running, you
don't need console access. Before the operating system is running,
having console access can be useful, but that should be rare.
>2) Some internal Ethernet switch that I have to learn to manage.
> We run a tight, integrated cloud of ethernet switches, centrally
> managed, with very differentiated per-port VLAN settings. We use
> Cisco Switches (Cat650x in the core, 2980 for aggregation).
> Each blade server comes with _some_ switch, all have their
> limitations, and would need careful integration, if possible
> at all.
>
>Both criteria are central for us: we build systems into racks, and
>LATER parcel them out to one or the other administrator, after setting
>them up (anew, if it's a project switch), and after putting them into
>the desired VLAN. Those other administrators work remotely, per default,
>and need individual access to a management console.
Why not use SSH or VNC?
> .... They are just
>developed, administered, and deployed by various small teams of people,
>each team responsible for the (web/db) applications on one or two
>groups of systems.
Seems like a good idea to me.
In summary: I think there are some applications where blades are
useful, but I'm not claiming they're perfect for everyone.
Regards,
Jan.
Peter da Silva
> x335, 2 x Xeon-2400, 1 x 36 GB SCSI drive, 1.5GB RAM =
> $9685 each.
>
> Cost for a 84 CPUs (42U) =
> 42 * $9685 =
> $406770
>
> HS20, 2 x Xeon-2400, 1 x 40GB IDE drive, 1.5GB RAM =
> $8985 each.
> Chassis, holds 14 blades (7RU) =
> $7919
>
> Cost for a 84 CPUs (21RU) =
> 3 * $7919 + 3 * 14 * $8985 =
> $401127
But you can fit at least 2x and sometimes 4x the drives in 1U
servers, and you can fit significantly more RAM (1.5G? That's a
game machine), I would be stunned if you couldn't get 4-way or
more boxes. Also, 1U servers are fungible... I can replace a 1U
IBM with a 1U Dell or a 1U Macintosh or a 1U Sparc.
Blades are basically targeted for an environment where you have to
run a separate system for each application instance. I can get a
1U PC and install FreeBSD on it, and split it into 10 virtual
systems running (as far as the customer is concerned) their own
copies of FreeBSD or Debian Linux.
--
`-_-' Ar rug tú barróg ar do mhactíre inniu?
'U` Peter da Silva
Peter da Silva
> On Wed, 19 Mar 2003 16:37:42 -0600, Peter da Silva <pe...@abbnm.com>
> wrote:
>>http://www.theinquirer.net/?article=7078
> Ummm, a *teraflop*? In a game console? Is this for real?
I don't see why not. What processor power would you need to do
photorealistic raytracing with a radiosity-driven illumination
model and realistic physics (including linear and angular
momentum, realistic hair and fur, wind, and friction) at 30 frames
per second at HDTV resolution?
That's the current "sound barrier" for games machines.
Peter da Silva
> Peter da Silva wrote:
>
>> Michael S wrote:
>>
>>> No. Desktops are driven by gaming. Games are hungry for CPU cycles.
>>> Pentium 4 delivers.
>> Multiprocessors deliver CPU cycles better,
> ? Define "better."
Image rendering is a highly parallelizable problem, and is the
core problem for games consoles. Everything else is secondary to
realism and frame rates.
>> if it wasn't for the Compaqtion,
> Good term! I'm going to use it. Is it original with you? (I like
> attribution.)
I don't recall hearing it from anyone else, but you might want to
check on google... I don't track these things myself.
Peter da Silva
> In article <b5ar6v$1fig$2...@jeeves.eng.abbnm.com>,
> Peter da Silva <pe...@abbnm.com> writes:
> |> Nick Maclaren wrote:
> |> > My prediction is that we shall see OEMs start to produce
> |> > blade-based desktops (i.e. the CPU board of a desktop will be just
> |> > a server blade), and that purpose-designed desktops could well be
> |> > quite rare in a couple of years (being replaced by laptops and
> |> > blade-based systems).
> |> One hopes that they're not quite as tightly integrated as blades.
> |> I would love a backplane based desktop rather than a motherboard
> |> based one, but having everything integrated with the processor
> |> card drops you right back in the same forklift-upgrade-only world.
> Given that a decreasing proportion of the market is to people (or
> even retailers) who can or will replace anything smaller than a
> major board, I am afraid that you are out of luck. Note that I am
> not saying that this is a desirable change - merely that it is the
> situation.
That is *why* I'm hoping they're not as tightly integrated as blades.
Bill Todd
"Peter da Silva" <pe...@abbnm.com> wrote in message
news:b5e316$2kfb$1...@jeeves.eng.abbnm.com...
> Greg Pfister wrote:
> > Peter da Silva wrote:
...
> >> if it wasn't for the Compaqtion,
>
> > Good term! I'm going to use it. Is it original with you? (I like
> > attribution.)
>
> I don't recall hearing it from anyone else, but you might want to
> check on google... I don't track these things myself.
It reminds me of some of Terry Shannon's (Shannon Knows DEC/Compaq/HPC/?)
quips, but I can't swear that I've heard this particular one from him.
- bill
Andy Glew
> end of the line for the Pentium 4 - not in 2003, obviously, but in
> the long term?
>
> If this is so, then the Prescott could well be Intel's last major
> success with a primarily desktop chip, and any future Pentium 4
> or successor would be squeezed out of the desktop market by the
> "lower" end in the same way that the RISC chips were by the x86
> range. My guess is that is what will happen, but I can't guess
> whether Intel will attempt to produce a successor design (and I
> don't just mean a new chip).
Yes, I can easily imagine this happening.
Banias may easily be more important to
Intel's future success than Pentium 4.
One of the reasons I left Intel was my feeling
that low power CPUs like Banias might be one
of Christenson's "disruptive technologies"
- and, by staying in Intel's Desktop Processor
Group in Oregon, I would be working on the
old, has-been, technology, while the guys
in Israel working on Banias (and the guys in
Arizona working on StrongARM) were where
the future lay.
---
I might have been willing to continue working on the
IMHO soon to be obsolete as a product line
desktop microprocessors - even if desktop PC
eventually dies out, the CPU microarchitectures developed
can still apply to where the market is - but I wasn't having fun
anymore. Another symptom of a mature product
line is that it gets tangled up in bureaucracy.
Ironically, I am not now working on low power
at AMD --- but I am at least having fun. AMD is a
lot more dynamic than Intel was when I left.
To be perfectly honest, AMD has much of the feeling
that the original P6 project did.
---
Anyway, back to topic:
I do not mean to say that performance no longer matters.
I think there is still demand for increased performance.
And that this demand will continue to need more advanced
CPU microarchitectures.
Just that, I don't think that performance outweighs all other
considerations. Performance at the cost of more power,
bigger fans, more expensive cooling systems, greater noise,
probably not.
Andy Glew
>
> Given that profiling have shown Intel that 30-50 stage pipelines can be
> a win, P4-style computing won't go away.
I guess I disappoint you too, Terje,
since I agree with Nick.
Depends on what you mean by P4-style computing.
I have many problems with the Willamette microarchitecture.
IMHO it lacks elegance, and I think that "elegance" is
a concept that evolved as a way of predicting "less likely
to have problems".
But I think that deep pipelines will remain in the
microarchitects' bag of tricks.
My problem is with the ever increasing power and cooling
demands of desktop microprocessors, and with their large
die sizes. I have no problem with large numbers of transistors
or pipestages, I'm just willing to wait a bit longer until they
can fit onto smaller die. My problem is with the "desktop"
PC form factor, that can no longer fit on top of a desk,
and which sometimes has trouble fitting into a tower
beside the desk.
Since I started in this industry, both chips and PC boxes
have grown physically bigger and more power hungry.
This is progress?
Remember when the usual thing that you marvelled at was
at how many devices fit onto a chip no bigger than
your thumbnail?
Maybe I'm just excessively in love with miniaturization.
Andy Glew
"Jakob Engblom"
> Nick Maclaren
> > Does anyone else think that Intel's Banias CPU could well spell the
> > end of the line for the Pentium 4 - not in 2003, obviously, but in
> > the long term?
>
> performance terms.
No. All it requires is that Banias outsells P4.
Which I doubt will happen - although I suspect
that Banias will take a big chunk of the market.
> Banias won't clock as highly due to its design it seems. 1.6 Ghz seems to
be
> a pretty tight max frequency for the current Banias, according to
> reports
Maybe Banias won't clock too fast. Not surprising.
Banias is descended from the P6 core - reworked a bit,
with better branch prediction, micro-op fusion.
But still a P6 at heart, a microarchitecture conceived
more than 10 years ago.
Of more concern to me, as an Intel competitor, is what
the Banias replacement may look like. If Banias shows
that there is a market for this sort of lower power chip,
then Intel Israel will probably have the go-ahead to
produce a microarchitecture designed more (rather than less)
from scratch to address this market.
Of course, that's what makes it fun, competition...
---
On the other hand, I would not be at all surprised to find out
that Intel management holds Banias back so as not to undermine
Pentium 4 sales. Just as Pentium 4 has probably been held
back so as not to undermine Itanium - well, not exactly sales,
but at least not to undermine the Itanium story.
I think that part of the reason for the emphasis on Centrino
rather than Banias is to attempt to differentiate Banias and P4
in the marketplace, to reduce the chance of a cheaper Banias
chip undermining Pentium 4 sales.
Although, it's not clear than Banias will be less profitable per chip
than Pentium 4.
Andy Glew
> Desktop->Workstation->Server.
In terms of market weight
Desktop 100->Workstation 10->Server 1
-> Laptop 20
(Sorry, I'm disconnected again,
or I'd look up current market shares,
both unit and $$$)
But laptops grow fast.
It will soon be
Desktop 100->Workstation 10->Server 1
-> Laptop 40
and, maybe soon
Laptop 50 -> Low-end desktop 60
-> Gamer desktop 40 -> Workstation 10 -> Server 1
I can easily imagine a scenario where it is
Laptop 50 + Low-end desktop 70 => Gamer desktop 30
->: Workstation 10 => Server 1
and, eventually, the middle being squeezed out.
Just as, now, workstations are not that big a consideration:
servers, and HPC (the stuff we used to call supercomputing)
are more influential. Reason: high end server and HPC
technology trickles down, while targetting the biggest market
maximizes $$$. The guys in the middle, while nice to sell
to, a good place to make profits, don't influence design that
much.
Stephen Fuld
"Andy Glew" <andy_gle...@yahoo.com> wrote in message
news:u0yea.35$1k2.4...@newssvr15.news.prodigy.com...
> > No. Economics of scale work in opposite direction.
> > Desktop->Workstation->Server.
>
> In terms of market weight
>
> Desktop 100->Workstation 10->Server 1
> -> Laptop 20
>
> (Sorry, I'm disconnected again,
> or I'd look up current market shares,
> both unit and $$$)
>
> But laptops grow fast.
>
> It will soon be
>
> Desktop 100->Workstation 10->Server 1
> -> Laptop 40
>
> and, maybe soon
>
> Laptop 50 -> Low-end desktop 60
> -> Gamer desktop 40 -> Workstation 10 -> Server 1
>
> I can easily imagine a scenario where it is
>
> Laptop 50 + Low-end desktop 70 => Gamer desktop 30
> ->: Workstation 10 => Server 1
>
> and, eventually, the middle being squeezed out.
Isn't there a bifurcation of the laptop market into the "desktop
replacement" segment and the "truely portable" segment? That may make some
difference as the demand for low power consumption for the desktop
replacement market may be considerably less.
Terje Mathisen
> Terje Mathisen <terje.m...@hda.hydro.com> writes:
>
>>Since the performance ratio between PIII and P4 class cpus vary a lot
>>between SPEC on one side and BAPCO (i.e. mostly Office) on the other
>>side, the longer pipeline in the P4 must cost more for those apps.
>>
>>Do you have any other explanation for the performance differences?
>
>
> SPEC is provided as source, and BapCo provided as binaries; the SPEC
> numbers are produced with Intel's C compiler, with Pentium 4 specific
> optimizations, many of the BapCo binaries are produced with other
> compilers, and not necessarily with optimzations for Pentium 4 (if the
> compiler version that was used for the binary already knew about the
> Pentium 4 at all).
Yes?
How do this disagree with my conclusion?
I claim that BapCo probably contains code with shorter average
instruction length between (mis-predicted) branches than SPEC.
It is quite possible that this is at least partly caused by using
different compilers and optimization settings on the two sets of source
code. I believe it is also caused by differences in coding style
(relatively clean C vs twisty little passages, all alike in C++).
The interesting point is still the huge difference in performance ratios!
Patrick Schaaf
>I went to the Australian IBM site (www.ibm.com.au) are did a comparison
>between the x335 (dual P4-2400) and the HS20 blade and chassis. I'm
>not advocating IBM; they just happened to be there.
Thanks for the comparison.
>So, the MIPS/$ is essentially identical. However, the
>blade approach was half the price in terms of rack units consumed.
>Where floorspace is expensive, this can add up.
Agreed. Blade servers appear to me to be the answer to the
crowded rack space in housing centers during the Internet boom.
I heard that the situation has cooled down, somewhat. On the
other hand, there should be enough crowded server racks at
SMEs, where I think blades will find their market, now.
>>Personal reasons why no blade I've seen appeals to me:
>>
>>1) System management console (extra ethernet) not per individual blade.
>> You can usually switch back and forth, but no concurrent operation
>> on different blade appears possible, and the security model is lacking.
>I don't see this as a big issue, but it might be important for you.
>Even 1RU servers typically have a switch where only one person has
>"console" access at a given time.
That's OK. One server, one Admin. With a blade system, you have
one person accessing and blocking 12 consoles at one time. Even
if the management console access is rare, when it is needed, it
is needed badly, and there's nothing more frustrating than "wait
until the other guy has fixed HIS box, before you can".
>My take on this is that once the operating system is running, you
>don't need console access. Before the operating system is running,
>having console access can be useful, but that should be rare.
It is rare. The consoles are primarily used during initial setup,
especially when the boot process does funny things, and they are
used later when access via the normal Ethernet connection is not
possible or stable (e.g. you are just reconfiguring the network).
Additionally, given the stability records of recent Linux kernels,
there's another feature of the Compaq RILO management board that
we like to use: they record the last 30 seconds' screen output
BEFORE the last machine reset. Thus, even after a panic-and-reboot,
the admin can still look at how the system crashed, even when the
syslogd has long been dead.
While use of the management cards is rare for each individual system,
with over 500 servers, you learn to love it.
>Why not use SSH or VNC?
We do use SSH for normal everyday administration.
>In summary: I think there are some applications where blades are
>useful, but I'm not claiming they're perfect for everyone.
Again, I can agree with that. The intent of my post was to show two
reasons why blade servers appear suboptimal for OUR setup.
We always hear that we are not the "typical customer". Makes me proud :)
best regards
Patrick
Andy Glew
> replacement" segment and the "truely portable" segment? That may make
some
> difference as the demand for low power consumption for the desktop
> replacement market may be considerably less.
Yes (I'm writing this on a desktop replacement.
I'd love to have a Tablet PC, but don't want to trade
off smaller screen for pen.)
Desktop replacement certainly means that giving up
too much performance for lower power is out. Ideally, you
want the desktop replacement to be just as high performance
as a desktop when plugged in.
But nevertheless, desktop replacements must dissipate less
power. It's harder to cool laptops. Until docking stations
start providing chilled water along with all the wires.
Greg Lindahl
Terje Mathisen <terje.m...@hda.hydro.com> wrote:
>Since the performance ratio between PIII and P4 class cpus vary a lot
>between SPEC on one side and BAPCO (i.e. mostly Office) on the other
>side, the longer pipeline in the P4 must cost more for those apps.
>
>Do you have any other explanation for the performance differences?
I never realized the only major difference between the PIII and P4 was
the pipeline length; I always thought that things like branch
prediction, differences in the OOO engine, memory bandwidth, memory
latency, and lots of other stuff was also important.
If you're going to look at ratios, you might want to look at the
scaling of the individual SPECcpu elements vs. BAPCO. Along the way
you'll probably observe that the individual SPECcpu components don't
have the same ratios either.
-- greg
Terje Mathisen
> In article <b5d4ls$k0q$1...@vkhdsu24.hda.hydro.com>,
> Terje Mathisen <terje.m...@hda.hydro.com> wrote:
>
>
>>Since the performance ratio between PIII and P4 class cpus vary a lot
>>between SPEC on one side and BAPCO (i.e. mostly Office) on the other
>>side, the longer pipeline in the P4 must cost more for those apps.
>>
>>Do you have any other explanation for the performance differences?
>
>
> I never realized the only major difference between the PIII and P4 was
> the pipeline length; I always thought that things like branch
Ouch, that's was a low hit, Greg. :-)
> prediction, differences in the OOO engine, memory bandwidth, memory
> latency, and lots of other stuff was also important.
Improved branch prediction was probably one of the key requirements
before P4 could look good in simulation, right?
I do agree that having most of the other, not pipeline length related,
improvements retrofitted to a P6 core, would make comparisons easier and
also more relevant. It seems like this is sort of what we're getting
now, anyway. :-)
Jan Mikkelsen
>Jan Mikkelsen wrote:
>> [ pricing on 1RU servers vs. blades from IBM ]
>
>But you can fit at least 2x and sometimes 4x the drives in 1U
>servers, and you can fit significantly more RAM (1.5G? That's a
>game machine), I would be stunned if you couldn't get 4-way or
>more boxes. Also, 1U servers are fungible... I can replace a 1U
>IBM with a 1U Dell or a 1U Macintosh or a 1U Sparc.
1.5GB is not the limit, it's just the amount I chose for pricing.
This particular model can go to 8GB RAM and can hold two drives, but I
don't see that as a big issue. With FreeBSD you have a practical
limit of around 3GB physical RAM in any case.
I don't understand the 4-way point. All the larger SMP systems I see
seem to reduce MIPS/RU density rather than increase it. As for a
4-way blade: why bother? I'd be happy with uniprocessor blades.
(Incidentally, I saw an 1RU machine somewhere with two uniprocessor
motherboards. Given the premium vendors are trying to charge for SMP,
it seems like a reasonable approach to me.)
As for vendor independence, yes, that is a blade drawback. However,
once you are filling racks with machines, the need to replace a single
machine with one from another vendor is less likely to come up.
>Blades are basically targeted for an environment where you have to
>run a separate system for each application instance. I can get a
>1U PC and install FreeBSD on it, and split it into 10 virtual
>systems running (as far as the customer is concerned) their own
>copies of FreeBSD or Debian Linux.
I don't know where they are targeted by the vendors, but I can
certainly see applications for blades which don't fit the model you
describe. The issues I am looking at have nothing to do with multiple
machine instances; the problem is just to apply CPU resources to
neatly contained transactional tasks.
I don't need a blade to have any local storage. Network boot or a
compact flash slot would also do. The applications are CPU bound
commercial applications where all the persistent state is stored on a
database server and/or in a recoverable message queue. Why waste
space and money on extra, empty, hot swap SCSI slots?
Regards,
Jan.
Jan Mikkelsen
>>blade approach was half the price in terms of rack units consumed.
>>Where floorspace is expensive, this can add up.
>
>Agreed. Blade servers appear to me to be the answer to the
>crowded rack space in housing centers during the Internet boom.
>I heard that the situation has cooled down, somewhat. On the
>other hand, there should be enough crowded server racks at
>SMEs, where I think blades will find their market, now.
The crowded rack problem is also getting smaller as machine just plain
get faster.
>>>Personal reasons why no blade I've seen appeals to me:
>>>
>>>1) System management console (extra ethernet) not per individual blade.
>>> You can usually switch back and forth, but no concurrent operation
>>> on different blade appears possible, and the security model is lacking.
>
>>I don't see this as a big issue, but it might be important for you.
>
>>Even 1RU servers typically have a switch where only one person has
>>"console" access at a given time.
>
>That's OK. One server, one Admin. With a blade system, you have
>one person accessing and blocking 12 consoles at one time. Even
>if the management console access is rare, when it is needed, it
>is needed badly, and there's nothing more frustrating than "wait
>until the other guy has fixed HIS box, before you can".
I was unclear, although I take your point. I meant that a set of 1RU
servers have a switch where only one console can be active at once.
For example, a set of the IBM 1RU servers I mentioned (x335) have
(essentially) an embedded KVM switch so that you just daisy chain them
and put your keyboard and screen at the end. Only one machines has a
physical console at any given time; this could be 1 console per 42
machines. I guess you at least have a choice about how many consoles
you want.
The other option is something like a serial console. See the
www.realweasel.com; they have my favourite product logo.
>>In summary: I think there are some applications where blades are
>>useful, but I'm not claiming they're perfect for everyone.
>
>Again, I can agree with that. The intent of my post was to show two
>reasons why blade servers appear suboptimal for OUR setup.
>
>We always hear that we are not the "typical customer". Makes me proud :)
I hear the same thing and I agree! Sometimes I think the typical
customer is one who has outsourced all decision making and has lost
control of his cheque book ...
Regards,
Jan.
Nick Maclaren
In article <LSxea.34$Se2.4...@newssvr15.news.prodigy.com>,
"Andy Glew" <andy_gle...@yahoo.com> writes:
|>
|> No. All it requires is that Banias outsells P4.
|> Which I doubt will happen - although I suspect
|> that Banias will take a big chunk of the market.
Yes. But, if that happens enough to shift people's viewpoints,
a Banias successor will probably outsell any Pentium 4 successor.
|> Of more concern to me, as an Intel competitor, is what
|> the Banias replacement may look like. If Banias shows
|> that there is a market for this sort of lower power chip,
|> then Intel Israel will probably have the go-ahead to
|> produce a microarchitecture designed more (rather than less)
|> from scratch to address this market.
|>
|> Of course, that's what makes it fun, competition...
Yes, indeed. And my interest is in systems that are less space
and power hungry - was it you who pointed out that "desktops" have
been growing so that they won't fit on desks any longer? This
affects many of us in HPC very badly, as we can get less into a
machine room.
|> On the other hand, I would not be at all surprised to find out
|> that Intel management holds Banias back so as not to undermine
|> Pentium 4 sales. Just as Pentium 4 has probably been held
|> back so as not to undermine Itanium - well, not exactly sales,
|> but at least not to undermine the Itanium story.
Agreed. But I don't think that Intel management will be able to
walk either tightrope for much longer. That approach works very
well when the competition is negligible, but tends to fall apart
in the face of a real threat. And both the laptop/desktop and
workstation/server boundaries are major targets of many important
competitors.
Nick Maclaren
In article <3e7ab92b$0$21717$9b62...@news.freenet.de>,
mailer...@bof.de (Patrick Schaaf) writes:
|>
|> >>Personal reasons why no blade I've seen appeals to me:
|> >>
|> >>1) System management console (extra ethernet) not per individual blade.
|> >> You can usually switch back and forth, but no concurrent operation
|> >> on different blade appears possible, and the security model is lacking.
|>
|> >I don't see this as a big issue, but it might be important for you.
|>
|> >Even 1RU servers typically have a switch where only one person has
|> >"console" access at a given time.
|>
|> That's OK. One server, one Admin. With a blade system, you have
|> one person accessing and blocking 12 consoles at one time. Even
|> if the management console access is rare, when it is needed, it
|> is needed badly, and there's nothing more frustrating than "wait
|> until the other guy has fixed HIS box, before you can".
In our case, we rarely have that problem, because we have so many
more nodes than people :-(
You are quite correct, though, and one of the more ghastly marketing
nonsenses is to claim that blade servers are a good way of merging
multiple, separate, independently managed systems into a single box.
That is complete nonsense. Where they score is with merging servers
that are managed as a cluster.
Jakob Engblom
> I'd be interested to see Pentium M SPEC benchmarks: I'm expecting
> something comparable to a 2400MHz Xeon.
>
Benchmarks are already up on the web. Not SPEC, but more real-world
benchmarks reported by PC enthusiast sites.
http://www.anandtech.com/mobile/showdoc.html?i=1800&p=13
And yes, for regular integer codes it can match the 2.4 Ghz P4. For
work that uses SSE2 and is compute bound, the P4 races far ahead.
Patrick Schaaf
>I was unclear, although I take your point. I meant that a set of 1RU
>servers have a switch where only one console can be active at once.
>For example, a set of the IBM 1RU servers I mentioned (x335) have
>(essentially) an embedded KVM switch so that you just daisy chain them
>and put your keyboard and screen at the end. Only one machines has a
>physical console at any given time; this could be 1 console per 42
>machines. I guess you at least have a choice about how many consoles
>you want.
We've seen those daisy chains in presentations, and avoided them
for the same reasons as we'd avoid either-or managed blades.
We incurr a significant per-system cost that way, I think the
RILO cards and the ethernet ports we plug them into, cost
something like $500 per system. But in our environment, that
is a sensible insurance cost.
best regards
Patrick
Jakob Engblom
Well, I believe that there is more of a continuum of
price-performance-battery life options. I have a rather portable
machine that does serve as a desktop replacement. But I still require
eight hours of battery life for working on trains and flights across
Europe and the Atlantic. But due to the nature of work, I need the
fastest CPU possible (doing computer simulation, which is inherently
100% serial).
And such a beast was possible to find; DELL latitude series answered
nicely (except for a lack of built-in WiFi, firewire and bluetooth).
So yes, something was sacrified to gain the form factor required.
A multipro laptop could solve some of my problems. But the biggest
performance killer is the slow disk...
/jakob
Patrick Schaaf
>You are quite correct, though, and one of the more ghastly marketing
>nonsenses is to claim that blade servers are a good way of merging
>multiple, separate, independently managed systems into a single box.
>That is complete nonsense. Where they score is with merging servers
>that are managed as a cluster.
I assume when you write cluster, you think upwards of 100 nodes,
with homogenous, centrally managed workload? I fully agree that
lades may make total sense.
Our average "cluster" has about 5 identical machines sharing
a certain workload (read: wasting their CPU cycles on the
same set of PHP scripts), and that's the "unit of administration".
I think the largest of the pack is 16 nodes wide.
For "fluffy resilience" reasons, we'd optimally want all nodes
of any of these small clusters to be in different racks, on
seperate aggregation switches, and those spread evenly over
the redundant core switches. General theme: minimize the effect
of any single failure. Five clusters a bit slower is OK, one of
them completely down is BAD.
That overall theme then forces us to have each blade in a single
blade server dedicated to a different task than each other blade
in the same server. And that maximizes the either-or administration
problem.
best regards
Patrick
Nick Maclaren
In article <3e7aecfd$0$5948$9b62...@news.freenet.de>, mailer...@bof.de (Patrick Schaaf) writes:
|> nm...@cus.cam.ac.uk (Nick Maclaren) writes:
|>
|> >You are quite correct, though, and one of the more ghastly marketing
|> >nonsenses is to claim that blade servers are a good way of merging
|> >multiple, separate, independently managed systems into a single box.
|> >That is complete nonsense. Where they score is with merging servers
|> >that are managed as a cluster.
|>
|> I assume when you write cluster, you think upwards of 100 nodes,
|> with homogenous, centrally managed workload? I fully agree that
|> lades may make total sense.
Not necessarily. The key is whether it is a 'unit' in any real
sense, rather than a bunch of separate machines in the same room
(or even rack).
|> Our average "cluster" has about 5 identical machines sharing
|> a certain workload (read: wasting their CPU cycles on the
|> same set of PHP scripts), and that's the "unit of administration".
|> I think the largest of the pack is 16 nodes wide.
That is a cluster, because of the homogeneity.
|> For "fluffy resilience" reasons, we'd optimally want all nodes
|> of any of these small clusters to be in different racks, on
|> seperate aggregation switches, and those spread evenly over
|> the redundant core switches. General theme: minimize the effect
|> of any single failure. Five clusters a bit slower is OK, one of
|> them completely down is BAD.
Interesting. A 2-D structure! And the clusters are across the
racks (or blade servers, if you have them) and not within them.
|> That overall theme then forces us to have each blade in a single
|> blade server dedicated to a different task than each other blade
|> in the same server. And that maximizes the either-or administration
|> problem.
I hadn't thought of your scenario, and I agree that blade servers
as currently constituted don't make much sense for you.
Rupert Pigott
[SNIP]
> Just as, now, workstations are not that big a consideration:
> servers, and HPC (the stuff we used to call supercomputing)
> are more influential. Reason: high end server and HPC
> technology trickles down, while targetting the biggest market
My immediate reaction to that is : NO WAY does HPC
technology trickle down ! On the other hand the low
end appears to catch up & eat the HPC lunch.
The low-end stuff is usually fabricated using very
different technology, and that definately influences
the architecture. Two obvious HPC technology Dodos
that come to mind are ECL and GaAs. I'm really
struggling to think of a HPC technology that crept
down, hmmm... Lemme see : Superscalar, Pipelining
erm... OK, maybe you have a point.
Perhaps my memory is too short or too uP focussed. :)
Cheers,
Rupert
Terje Mathisen
>> Nick, you disappoint me.
>>
>> Given that profiling have shown Intel that 30-50 stage pipelines can be
>> a win, P4-style computing won't go away.
>
> I guess I disappoint you too, Terje,
> since I agree with Nick.
Ouch. I'm out-numbered, outclassed, and outvoted.
If this was chess, it would be a good time to suggest a draw. :-)
> Depends on what you mean by P4-style computing.
>
> I have many problems with the Willamette microarchitecture.
> IMHO it lacks elegance, and I think that "elegance" is
> a concept that evolved as a way of predicting "less likely
> to have problems".
Personally I started out by disliking the P4, however, since it does
give those great SPEC numbers, I've learned to accept it. With the
recent post about theoretically optimum pipeline lengths of 30-50
stages, I actually started to believe in it as well.
The stuff I do like is: Better bandwidth, sane fp (SSE2), good performance.
As I've posted previously, I hate the warts that it currently have, in
the form of some opcodes that didn't make it into the fast core, and as
a consequence have become a _lot_ slower in both relative and absolute
terms. However, this could mostly be fixed after another shrink or two.
> Since I started in this industry, both chips and PC boxes
> have grown physically bigger and more power hungry.
> This is progress?
My boxes have grown consistently smaller over the same period, the
original PC/XT/AT were _big_ and extremely heavy.
> Remember when the usual thing that you marvelled at was
> at how many devices fit onto a chip no bigger than
> your thumbnail?
>
> Maybe I'm just excessively in love with miniaturization.
I want an SMT/SMP laptop, so I wouldn't need a big hulking workstation
just to develop & test multi-threaded code.
I would also like it to not get so hot that it can cause blisters in
awkward places. (True story: A swede got a 2 cm degree two burn on a
very delicate part of his anatomy.)
Nick Maclaren
In article <10482493...@saucer.planet.gong>,
"Rupert Pigott" <r...@dark-try-removing-this-boong.demon.co.uk> writes:
|> "Andy Glew" <andy_gle...@yahoo.com> wrote in message
|> news:u0yea.35$1k2.4...@newssvr15.news.prodigy.com...
|>
|> > servers, and HPC (the stuff we used to call supercomputing)
|> > are more influential. Reason: high end server and HPC
|> > technology trickles down, while targetting the biggest market
|>
|> My immediate reaction to that is : NO WAY does HPC
|> technology trickle down ! On the other hand the low
|> end appears to catch up & eat the HPC lunch.
|>
|> The low-end stuff is usually fabricated using very
|> different technology, and that definately influences
|> the architecture. Two obvious HPC technology Dodos
|> that come to mind are ECL and GaAs. I'm really
|> struggling to think of a HPC technology that crept
|> down, hmmm... Lemme see : Superscalar, Pipelining
|> erm... OK, maybe you have a point.
|>
|> Perhaps my memory is too short or too uP focussed. :)
Yes, that's it :-)
A lot of the technology does trickle down, but quite a lot
peters out before it gets down to the bottom. Many aspects
are relevant to servers, but not workstations - e.g. much of
the RAS, high-performance interconnects, management etc.
Others are relevant to high-end workstations, but not really
to the ordinary desktop and below. And so on.
Exactly the same applies to the stuff coming up from below,
though all details are different. I agree with Andy Glew that
the days that workstations led the innovation are over, at least
for the forseeable future.
Patrick Schaaf
>|> For "fluffy resilience" reasons, we'd optimally want all nodes
>|> of any of these small clusters to be in different racks, on
>|> seperate aggregation switches, and those spread evenly over
>|> the redundant core switches. General theme: minimize the effect
>|> of any single failure. Five clusters a bit slower is OK, one of
>|> them completely down is BAD.
>Interesting. A 2-D structure! And the clusters are across the
>racks (or blade servers, if you have them) and not within them.
We don't really slavishly enforce it, except for select systems.
It happens to just work as each cluster grows over time, system
by system. We basically filled one rack after the other with
incremental updates to whatever cluster (web farm) happened
to need another machine, along with the first, or first two
systems of new farms. This way, a good distribution occurs
without much planning ahead. As a side benefit, it keeps actual
work in the machine room localized, which is good for overall
stability.
Occasionally, visitors wonder how we can sanely manage a four
machine cluster, when it's spread over three rooms. We smile,
and point to an Ethernet switch.
NOTE: we enjoy the benefit of rather underutilized Ethernet
ports on each machine. Your bandwidth requirements will vary,
and then taking our approach could be very silly.
best regards
Patrick
Dan Pop
>I want an SMT/SMP laptop, so I wouldn't need a big hulking workstation
>just to develop & test multi-threaded code.
>
>I would also like it to not get so hot that it can cause blisters in
>awkward places. (True story: A swede got a 2 cm degree two burn on a
>very delicate part of his anatomy.)
From a *properly* used laptop?!?
Dan
--
Dan Pop
DESY Zeuthen, RZ group
Email: Dan...@ifh.de
Keith R. Williams
> In <b5f0h8$s8d$1...@vkhdsu24.hda.hydro.com> Terje Mathisen <terje.m...@hda.hydro.com> writes:
>
> >I want an SMT/SMP laptop, so I wouldn't need a big hulking workstation
> >just to develop & test multi-threaded code.
> >
> >I would also like it to not get so hot that it can cause blisters in
> >awkward places. (True story: A swede got a 2 cm degree two burn on a
> >very delicate part of his anatomy.)
>
> From a *properly* used laptop?!?
Well, FWIW: http://www.mydr.com.au/default.asp?article=3825
--
Keith
Terje Mathisen
> In <b5f0h8$s8d$1...@vkhdsu24.hda.hydro.com> Terje Mathisen <terje.m...@hda.hydro.com> writes:
>
>
>>I want an SMT/SMP laptop, so I wouldn't need a big hulking workstation
>>just to develop & test multi-threaded code.
>>
>>I would also like it to not get so hot that it can cause blisters in
>>awkward places. (True story: A swede got a 2 cm degree two burn on a
>>very delicate part of his anatomy.)
>
>
> From a *properly* used laptop?!?
According to reports, he was wearing both underwear and shorts at the
time, so yes.
Greg Pfister
[snip]
>>>if it wasn't for the Compaqtion,
>
>>Good term! I'm going to use it. Is it original with you? (I like
>>attribution.)
>
> I don't recall hearing it from anyone else, but you might want to
> check on google... I don't track these things myself.
Good advice. I found refs going back to 2001 of Terry Shannon using it
in "Shannon Knows Compaq." I'd guess he's the author, but can ask him.
--
Greg Pfister
Greg Pfister
> Greg Pfister wrote:
>>Peter da Silva wrote:
>>>Michael S wrote:
>>>
>>>>No. Desktops are driven by gaming. Games are hungry for CPU cycles.
>>>>Pentium 4 delivers.
>
>>>Multiprocessors deliver CPU cycles better,
>
>>? Define "better."
>
> Image rendering is a highly parallelizable problem, and is the
> core problem for games consoles. Everything else is secondary to
> realism and frame rates.
OK, define "better" for image rendering. As you say, it's highly
parallelizable. As a result, it's a favorite demo for clusters of
systems that shows immense efficiency on those systems.
I suspect it's actually memory bandwidth limited, so adding processors
isn't going to help. But adding additional memory systems along with
the processors can help.
Cluster on a desktop? I admit, not likely at this time.
--
Greg Pfister
Tim McCaffrey
an...@mips.complang.tuwien.ac.at says...
>
>Terje Mathisen <terje.m...@hda.hydro.com> writes:
>>Since the performance ratio between PIII and P4 class cpus vary a lot
>>between SPEC on one side and BAPCO (i.e. mostly Office) on the other
>>side, the longer pipeline in the P4 must cost more for those apps.
>>
>>Do you have any other explanation for the performance differences?
>
>numbers are produced with Intel's C compiler, with Pentium 4 specific
>optimizations, many of the BapCo binaries are produced with other
>compilers, and not necessarily with optimzations for Pentium 4 (if the
>compiler version that was used for the binary already knew about the
>Pentium 4 at all).
>
P III has an instruction cache, the P4 has a trace cache.
P III has an AGU, the P4 does not.
P III has a bigger data cache.
- Tim
Hans de Vries
Well, the king is dead, long live the king (P5) they say.
About Prescott:
I saw on aces-hardware (CEBIT report) that the Intel's German Press man
said that Hyperthreading was improved by reducing cache aliasing.
I guess this has to do with the 64k aliasing issue of the L1 data cache.
Probably caused by the Way Predicting algorithm that I think should
look like this:
L1 DCache = 8kB = 4 ways of 2kB = 4 ways of 32 cache lines of 64 byte.
Address bits 6..10 (sum-indexed) are used to index into the 32 cache-lines.
Address bits 15..11 are used to for way-prediction:
The Tag bits 15..11 of all 4 ways are compared with address bits 15..11
The Way whose Tag matches is selected as the predicted one
(Like sum-indexed bit 6..10 one can use a CSA for to bypass the actual
address adder for bits 15:11)
But, all four Tags >must< be different for the predictor to work, thus...
A new cache line must replace any existing cache line with identical tag.
The 4-way mechanism is thus limited in the sense that it works like
a 1-way (direct mapped cache) for cache-lines with identical bits 15:11
and hence the 64kB aliasing problem. It gets bad with hyperthreading
(Imagine 2 identical threads running at the same time...)
Using more address bits may be a problem because of the 16 bit slices of
the Rapid Execution Engine. My guess is that the Thread ID is atached to
the Tag to solve the problem for Hyperthreading. At the other side,
Prescott 16 kB DCache reduces the available Tag address bits to 4.
I thought you might be interested because you did work on the P4 DCache...
Regards, Hans de Vries
Dan Pop
Thanks. Too bad they didn't mention the laptop brand...
Nick Maclaren
Brand being the operative word.
Paul DeMone
Terje Mathisen wrote:
>
> > From a *properly* used laptop?!?
>
> According to reports, he was wearing both underwear and shorts at the
> time, so yes.
You think he would claim otherwise? Ask any doctor who has worked
emergency in major center for a few years about some of the unusual
"predicaments" they occasionally encounter and the even more unlikely
explanation from the afflicted.
--
Paul W. DeMone The 801 experiment SPARCed an ARMs race of EPIC
Kanata, Ontario proportions to put more PRECISION and POWER into
pde...@igs.net architectures with MIPSed results but ALPHA's well
that ends well.
Robin KAY
See http://www.theregister.co.uk/content/54/28245.html
*Reg Reader Peter Dudek of Switzerland writes:
Just thought I'd mention that by doing a google search with "Do not
allow your portable computer to operate with the base resting directly
on exposed skin. With extended operation, heat can potentially build up
in the base. Allowing sustained contact with the skin could cause
discomfort or, eventually, a burn" as the search string, I believe,
reveals the mystery laptop brand.
It's a Dell Latitude. The search pulls up a support website from
support.jp.dell.com, that has, essentially, the exact same warning in the
"User's Guide".
--
Wishing you good fortune,
--Robin Kay-- (komadori)
Dennis M O'Connor
> Since I started in this industry, both chips and PC boxes
> have grown physically bigger and more power hungry.
> This is progress?
The Shuttle SFF boxes can run a 3.06GHz Pentium 4
processor and a high-power video card, while taking
up a lot less space than the original IBM PC. And the
micro-ATX Enlight case that my wife's P4 kitchen
computer is built inside is pretty small too.
So bigger boxes, if it even is a trend,
isn't being driven by the processors.
Processors are bigger, sure, but a the same time,
they are cheaper. Remember the price at launch
of the 80386 ? Now factor in inflation.
Frankly the processor could be the size of a silver
dollar and who would care, if the system it is in
costs less and performs better ?
So, yes, it is progress.
--
Dennis M. O'Connor dm...@primenet.com
"We don't become a rabid dog to destroy a rabid dog,
but we do get a bit rude."
Del Cecchi
area, on something hot until you got burned? Or held your finger under the hot
water faucet as the water went full hot?
Fails the plausibility test. I am not worried about my laptop burning me
without me noticing. Sorry.
|>
--
Del Cecchi
cec...@us.ibm.com
Personal Opinions Only
Robert Myers
wrote:
>
>Aside from the technological argument, consider the process of natural
>selection. If Banias couldn't offer performance comparable or better than
>Pentium 4M in a power constrained target platform then it wouldn't have
>been developed and brought to market. OTOH, if the P4 microarchitecture
>can't provide better performance for desktop form factor than a short pipe
>design Intel could have succeeded the P6 core with a Banias style design
>and saved itself the expense of building a good fraction of its 0.13 um fab
>capacity.
The fact is that Banias wasn't ready, and AMD was winning the
performance sweepstakes. Maybe before they die, the Intel executives
involved in inflicting the P4 on the world will admit to what really
happened.
RM
Keith R. Williams
cec...@signa.rchland.ibm.com says...
> In article <3E7B7685...@igs.net>,
> Paul DeMone <pde...@igs.net> writes:
> |>
> |>
> |> Terje Mathisen wrote:
> |>
> |> >
> |> > > From a *properly* used laptop?!?
> |> >
> |> > According to reports, he was wearing both underwear and shorts at the
> |> > time, so yes.
> |>
> |> You think he would claim otherwise? Ask any doctor who has worked
> |> emergency in major center for a few years about some of the unusual
> |> "predicaments" they occasionally encounter and the even more unlikely
> |> explanation from the afflicted.
> |>
> |> --
> |> Paul W. DeMone The 801 experiment SPARCed an ARMs race of EPIC
> |> Kanata, Ontario proportions to put more PRECISION and POWER into
> |> pde...@igs.net architectures with MIPSed results but ALPHA's well
> |> that ends well.
> |>
> How drunk was he? Have you ever tried to hold your hand, much less a sensitive
> area, on something hot until you got burned? Or held your finger under the hot
> water faucet as the water went full hot?
You're not suggesting intoxicating beverages are consumed by
geeks with laptops, surely! Shudder! Where was the P4
throttling?
> Fails the plausibility test. I am not worried about my laptop burning me
> without me noticing. Sorry.
Well, my antique 760EL did give a more than warm feeling when I
used it on my lap some years ago. I don't even go there with the
A21p. I like my privates as they are (and my adult beverages
cold ;-).
--
Keith
Robert Myers
<andy_gle...@yahoo.com> wrote:
>
>To be perfectly honest, AMD has much of the feeling
>that the original P6 project did.
>
If AMD's fundamentals didn't look so dicey, I'd be tempted to go out
and buy some stock on the strength of that.
RM
Peter da Silva
> 1.5GB is not the limit, it's just the amount I chose for pricing.
> This particular model can go to 8GB RAM and can hold two drives, but I
> don't see that as a big issue. With FreeBSD you have a practical
> limit of around 3GB physical RAM in any case.
True, if I was going to put more than 2GB in a 1U server I'd have
it running Tru64. I just brought up FreeBSD's jails as an example
of just how much partitioning you can get in a PC operating
system. The point is, if your OS doesn't get in your way there's
lots better mechanisms for running multiple application instances
than dedicating a CPU for each one.
> As for vendor independence, yes, that is a blade drawback. However,
> once you are filling racks with machines, the need to replace a single
> machine with one from another vendor is less likely to come up.
After a couple of years, even if you go back to the same vendor
you're finding yourself making a forklift upgrade regardless.
Vendor independence is useful even if you only have one vendor.
> I don't know where they are targeted by the vendors, but I can
> certainly see applications for blades which don't fit the model you
> describe. The issues I am looking at have nothing to do with multiple
> machine instances; the problem is just to apply CPU resources to
> neatly contained transactional tasks.
As you note, Blades aren't really that great a fit either. They
may be the least worst for that application, but if that was what
they were for they'd *certainly* be diskless.
--
`-_-' Ar rug tú barróg ar do mhactíre inniu?
'U` Peter da Silva
Peter da Silva
> I suspect it's actually memory bandwidth limited, so adding processors
> isn't going to help. But adding additional memory systems along with the
> processors can help.
How much memory would you need to give each processor a copy of
the polygons for the frame being rendered? Remembering that
commodity DRAM is currently selling around 10c a megabyte (I got 1
gigabyte of SDRAM for the wife for under $100).
Heck, you could probably make do just treating the L2 cache as
local RAM. :)
Nick Maclaren
Del Cecchi <cec...@us.ibm.com> wrote:
>|>
>How drunk was he? Have you ever tried to hold your hand, much less a sensitive
>area, on something hot until you got burned? Or held your finger under the hot
>water faucet as the water went full hot?
>
>Fails the plausibility test. I am not worried about my laptop burning me
>without me noticing. Sorry.
Perhaps you should be. That is not the mechanism that causes such
burns.
In the case of touching something hot with an unrestricted blood
supply, your pain receptors react (after a short delay) well below
the level at which serious tissue damage occurs.
In the case of something just above warm, especially when combined
with restricted blood supply (e.g. due to compression), this is not
always the case. The laptop probably reached only 42-44 Celsius.
This has never happened to me, but it is a known problem.
Jakob Engblom
> just to develop & test multi-threaded code.
>
Off-topic: try simulating that multithreaded machine instead, if it is
just for testing purposes. I do that all the time on my laptop (for
demos).
/jakob