68030/68040 questions
Warren Burnett
I am considering buying an accelerator for my 2000 but have not yet decided
what to get. Can anyone tell me what the differences are between the 68030
and 68040? I know that the 040 is supposed to be about twice as fast as an
030 at the same clock rate. I know that the 040 has an MMU and FPU on chip
as well as some cache, but I have heard that the 040's MMU and FPU are
subsets of the 030's MMU and 68882. I have also heard that the 040's FPU
is actually slower than the 68882 on certain problems. Thanks for any info.
Warren Burnett
war...@cup.hp.com
Michael Sinz
Well, this is what I have seen with the 68040 vs 68030.
The 68040, in an equally well designed system, is about 3 times as fast
as a 68030 at the same clock rate. (This depends on the software, I have seen
it as much as 500% and as low as 85% the speed of a 68030)
The 68040 has much more complex caches on the chip. This can cause bad
software to fail more easily than on the 68030. The 68040 is so fast that
some programs just could not deal with the preformance of the system.
The 68040 MMU is actually two MMUs and they are completely different from
the 68030. (Well, some things are the same) They were designed for
performance but, IMHO, have one flaw in that they don't have early page
termination.
The 68040 FPU is *very* fast. It does 80-bit extended IEEE multiply with
normalization and rounding in 5 cycles! (Yes, that is 5 cycles! The
68000 takes 4 cycles just to clear a register...) Other floating point
operations are also of that speed level. The FPU, being on chip, has
removed the interchip delay that external FPUs cause. However, since it
is on the main chip and they put so many transistors into the basic
instructions, the complex instructions are no longer done in hardware.
These instructions include all of the transindental functions such as
fsin, ftan, etc. These instructions are, in most systems, emulated in
software by the illigal instruction handler routines. While the 68040
can calculate SIN(x) very fast (faster than the 68881 at the same clock)
the overhead of also doing the exceptions makes it somewhat slower than
the 68881. However, as software starts to take advantage of the 68040
they will rely on the hardware fsin instruction less and less and implement
the SIN() function (and the others) as function calls.
The 68040 has a number of other features such as taken conditional branches
being faster than untaken branches (most loops only exit once while
the loop many times ;^) and instruction overlap/pipelining and register
scoreboarding for parallel execution.
/----------------------------------------------------------------------\
| /// Michael Sinz - Senior Amiga Software Engineer |
| /// Operating System Development Group |
| /// BIX: msinz UUNET: rutgers!cbmvax!mks |
|\\\/// "A master's secrets are only as good as the |
| \XX/ master's ability to explain them to others" |
\----------------------------------------------------------------------/
Don Faulkner
I heard the 68040 has a vector register. Does it? And if so, how does it
compare to the ones in larger computers, such as Crays?
Later,
Don
--
*******************************************************************************
| Donald Faulkner | Only on // /\ _ _ __ |
| de...@engr.uark.edu | // / \ |/ \_/ \ | / \ /\ |
| dfau...@uafhp.uark.edu | \\ // /----\ | | | | | _ /--\ |
| u7...@nes.nersc.gov | \X/ / \ | | | | \__/ / \ |
*******************************************************************************
Michael Sinz
[lots of quoted stuff deleted...]
>I heard the 68040 has a vector register. Does it? And if so, how does it
>compare to the ones in larger computers, such as Crays?
I think you are getting two terms confused here. The "Vector Base Register"
on the 68040 is just like that on the 68030, 020, and 010. It is a simple
offset register which is used to find the interrupt vectors and thus lets you
move the vectors from location $0 to whereever you want.
The Cray has a "Vector Processor" in the math term. That is, runs its math
operations through this array of processing hardware that makes the math go
much faster. The 68040 is not quite so complex but it can do multiple things
such as an addition and a multiplication at the same time as long as one does
not require the results of the other. In general, the operations are so fast
that it is hard to get much overlap. (With a 5-cycle 80-bit floating point
normallized multiply, you can see why...)
/----------------------------------------------------------------------\
| /// Michael Sinz - Senior Amiga Software Engineer |
| /// Operating System Development Group |
| /// BIX: msinz UUNET: rutgers!cbmvax!mks |
|\\\/// |
| \XX/ "I think not." said Ren'e Descartes, then he vanished. |
\----------------------------------------------------------------------/
Warren Burnett
Thanks to all who responded. After reading all the responses I think I
am going with an 040 board. They seem to offer much more bang for about
the same bucks as a 50 Mhz 030, and the limitations on the MMU and FPU should
not have much of an impact on the kind of work I do. Thanks again.
Warren Burnett
war...@cup.hp.com
Simon Raybould
So how about some suggestions for the best 40 card to get, I've been
considering getting the Fusion 40 card running at 50Mhz.
--
Simon J Raybould (s...@fulcrum.bt.co.uk) // {o.o}
"Only joking ... Or am I ?" - Vic Reeves \X/AMIGA \-/
===========================================================================
Fulcrum communications L.T.D., Fordrough Lane, Birmingham, B9 5LD, ENGLAND.
Dave Haynie
>So how about some suggestions for the best 40 card to get, I've been
>considering getting the Fusion 40 card running at 50Mhz.
***MARKETING HYPE ALERT***
Sorry, that sentence set off my bullshit detector. True, the Fusion 40 card
has a 50MHz clock, but its a 25MHz '040. All '040s run from a dual clock;
the "rating" of the part (eg, 25MHz or 33MHz) is the speed of the bus clock,
which determines how fast the CPU accesses memory, etc. This number is
comparable to the speed rating of '030s w.r.t. the memory speed requirements,
etc. The other clock, in this case the 50MHz clock, is a double speed clock
the processor clock, used for internal operations (a 33MHz '040 uses a 66MHz
PCLK).
--
Dave Haynie Commodore-Amiga (Amiga 3000) "The Crew That Never Rests"
{uunet|pyramid|rutgers}!cbmvax!daveh BIX: hazy
"I used to be disgusted, now I try to be amused" - Elvis Costello
Michael Sinz
>
>So how about some suggestions for the best 40 card to get, I've been
>considering getting the Fusion 40 card running at 50Mhz.
You may need to wait a while for that since there are not 50MHz 68040
chips. (Well, unless you want to call the 25MHz chip a 50MHz chip
since the input clock is 50MHz, but the bus clock is only 25MHz.)
/----------------------------------------------------------------------\
| /// Michael Sinz - Senior Amiga Software Engineer |
| /// Operating System Development Group |
| /// BIX: msinz UUNET: rutgers!cbmvax!mks |
David Tiberio
>considering getting the Fusion 40 card running at 50Mhz.
I thought the 68040 33mhz was just released last week! They boosted it
to 50 already? Lets see, probably about 45 mips...
--
David Tiberio SUNY Stony Brook 2-3662 AMIGA DDD-MEN
"Why do we have to go and take the same each day. Life is it what it is."
People don't change. You just get to know them better. -DT
Liverpool, New York...soon to be in 3D...starting with Heid's
William J. Coldwell
>considering getting the Fusion 40 card running at 50Mhz.
I think you'll be waiting a long time for a 50MHz 040... you might as
well be waiting for an 060. ;-)
Don't be fooled by a marketing pitch that claims that they have an 040 running
at 50MHz. They are in fact 25MHz externally, and doubles that speed internally
for the 040 pipeline. Some manufacturers are providing a 50MHz crystal and
dividing by 2 to get the 25MHz clock, leading to the illusion of an 040 being
50MHz (especially when they have a heat sink over the part number).
I won't go into it about manufacturers that are overrating parts and making
people believe there are 28MHz 040s in existance...
BTW - The 33MHz 040 is becoming more and more available, so expect to see
them in quantities on the Amiga RSN.
>Simon J Raybould (s...@fulcrum.bt.co.uk) // {o.o}
--
William J. Coldwell bi...@cryo.rain.com CryoCafe: (503)257-4823
"Amiga Attitude Adjuster" uunet!m2xenix!cryo!billc V.32/V.42bis/MNP5
Cryogenic Software Portal: Cryogenic Fish CDROM, UUCP
Warning: Potential Mood Swing Possible.
greg schaem
My question is: 040 50 mhz internaly with 25mhz bus
030 XX mhz internaly with 25mhz bus...
Do the 040 run 2 time faster in the cache than the 030?!?!
I dont see how they can get 4 time the speed if the 040 dont run
REALLY at 50mhz.And if the bus was 20mhz would it make a diference
in a A3000 since there is wait state a 25mhz?
I'm puzzled.
Stephan.
greg schaem
Well, would it make a diference if the 040 baord had a 50mhz
bus timming in an A3000? already my 25mhz 030 has waitstate
what good will it do to run at 50 or 100mhz if the ram dont follow?!
Is this ilogic?
Stephan.
Stephen Menzies
huh? what's this story about RCS (fusion40) and a market pitch claiming
a 50Mhz 040. I have an Amigaworld (nov91,pp63) infront of me right now.
The RCS ad claims : Motorola MC68040 microprocessor at 25MHhz. No
mention of 50mhz *anywhere*.
I also have a recently picked up RCS flyer (straight from their offices)
and it says "25mhz". No mention of 50mhz there either.
--stephen
--
Stephen Menzies
#Internet: men...@CAM.ORG
#Fidonet : Stephen Menzies @ 1:167/265
Dave Haynie
> My question is: 040 50 mhz internaly with 25mhz bus
> 030 XX mhz internaly with 25mhz bus...
XX = 25
>Do the 040 run 2 time faster in the cache than the 030?!?!
More like 3x-4x as the same bus clock speed, depending on the operation. The
68040 is quite a bit a bit more than a 68030 with a double-speed processor
clock. The '040 has two 4K 4-set associative caches vs. the '030's 256 byte
direct mapped caches, and the '040 caches hit twice as fast. The some of the
most common and simplest 680x0 instructions execute much faster on the '040
than on other 680x0 systems, as they're hardwired rather than microcoded.
Also, the '040's deeper pipeline allows more things to happen in parallel
and/or without delay than on an '030.
Steve Koren
greg schaem
right?
50mhz internal, cache 2 time faster + hardwired instructions...
2 time for clock speed, 2 time for chache hit x time for instruction
timing.So why do real life benchmark say the 040 run only 2 time faster?
I cant imagine memory downgrading so mutch. or is it the memory.
Stephan.
greg schaem
What do that have to do with my question about
A3000 with a 040 having a 50mhz bus speed.
The amiga memory run at 1 wait state with a 25mhz
bus speed, can 2 wait state be better? And dont
the internal cache use the other clock.
Can a 040 with a 50mhz bus clock vs 25mhz do any
good in a A3000?
Stephan.
Steve Koren
> > That's what caches are for. -Steve
> What do that have to do with my question about
> A3000 with a 040 having a 50mhz bus speed.
Well, your original question was, as I understood it, asking
how speeding up the clock running the 680x0 would help if the
memory did not get faster also. If the chip has internal
caches, some access to external RAM do not actually have to
go out to RAM at all. The bigger the cache, the less often
you should have to access external RAM. As long as you have
some internal cache, increasing the speed of your CPU will
still have an effect, even if your external RAM access can't
keep up.
- steve
Dan Green
SR> considering getting the Fusion 40 card running at 50Mhz.
Wait a minute! A Fusion 40 card running at 50MHz? Have I missed something
here or what? There aren't any 68040 processors running at 50MHz (as far as
I know!). Well, anyway, there are 2 -040 card which you might be interested
in: first of all the PP&S Mercury card, which runs at 28MHz (with easy
upgrade to 33MHz when availible) and then it's GVP's G-FORCE 040 card, also
running at 28MHz to date. The Mercury can hold up to 32Meg RAM while
G-FORCE supports up to 8Meg (or 16Meg?), I think that both of them supports
40ns RAM chips and 1Meg of SRAM. But these are the two leading cards at the
moment, Fusion 40 is an old one, but it's still an option (but you can
forget about the 50MHz version :-)).
wr.
_________________________
/ Dan Green / Sprintsoft /
/________________________/
Dan_Green%ca...@sweden.eu.net
-- Via DLG Pro v0.985b
bu...@vaxc.cc.monash.edu.au
So where are you filling your caches *from*? And where are your
cache writebacks going *to*? To main memory of course. Unless '040
boards start shipping with large external caches to augment the two 4K
caches on the 040 chip, you're going to have to use main memory sooner or
later.
Bu...@vaxc.cc.monash.edu.au OR bu...@monu1.cc.monash.oz
Alias: Gareth Bull, The Opal Dragon
DOD# 251 '84 VF 750 Closet Ducatisti
Huet dich bauer, ich komm! - Watch out farmer, I come!
Martin Hunt
> In article <2436...@hpfcdc.fc.hp.com>, ko...@hpfcdc.fc.hp.com (Steve Koren) writes:
> > sch...@sgi.com (greg schaem) writes:
> >
> >> Well, would it make a diference if the 040 baord had a 50mhz
> >> bus timming in an A3000? already my 25mhz 030 has waitstate
> >> what good will it do to run at 50 or 100mhz if the ram dont follow?!
> >> Is this ilogic?
> >
> > That's what caches are for.
>
> So where are you filling your caches *from*? And where are your
> cache writebacks going *to*? To main memory of course. Unless '040
> boards start shipping with large external caches to augment the two 4K
> caches on the 040 chip, you're going to have to use main memory sooner or
> later.
>
if you can do most of your reads and writes to/from cache, then you still
get a major speed increase, right?
--
-------------------------------------------------------------------------------
Martin Hunt mar...@cac.washington.edu
Networks and Distributed Computing University of Washington
Dave Haynie
>
> > That's what caches are for. -Steve
>What do that have to do with my question about
>A3000 with a 040 having a 50mhz bus speed.
Everything.
Assume you could get a 50/100MHz '040 (you can't, the fastest available is
the 33/66MHz part). It takes just as long as the 25MHz part to read or write
external RAM. Anything in cache, and any internal operation as well, goes
twice as fast. If you consider that this hit rate on an '040 cache is way
up there (I don't know offhand -- 80%? 90%? 95%? Mike?), you'll find that a
good portion of the time the faster processor would indeed go faster. Since
it'll never go slower, you'll get an improvement. Sure, it may not be 100%,
but 70% is nice.
Steve Koren
> >> bus timming in an A3000? already my 25mhz 030 has waitstate
> >> what good will it do to run at 50 or 100mhz if the ram dont follow?!
> >> Is this ilogic?
> >
> So where are you filling your caches *from*? And where are your
> cache writebacks going *to*? To main memory of course. Unless '040
Yes, of course from main memory. However, the point is that you don't need
to go to main memory nearly as *often*. The whole point of a cache is that
it is faster than main memory. If, like the first poster theorized,
there was no speed increase available from using caches, they wouldn't
bother to put them on the chips. They make a huge difference. If I
turn off both data and address caches on my system, it runs roughly 5X
slower. The bigger the difference in speed between your CPU and your
memory, the more caches help. That's why they exist. In designing a
computer system you usually have a small amount of really fast but expensive
storage (registers), a somewhat larger amount of less expensive but
still pretty fast storage (cache), etc, going through memory, disk
space, off-line storage. If having the faster storage didn't help,
we wouldn't have them and we'd be executing directly off backup tapes!
(If you still don't believe this, email me and I'll send a more detailed
explanation of how caches work).
- steve
Stephen Menzies
>[lines deleted]
>moment, Fusion 40 is an old one, but it's still an option (but you can
>forget about the 50MHz version :-)).
>wr.
> _________________________
> / Dan Green / Sprintsoft /
>/________________________/
>Dan_Green%ca...@sweden.eu.net
>-- Via DLG Pro v0.985b
The Fusion40 is also now 28mhz with upto 32megs onboard. I believe it's
the least expensive board as far as the US goes and here in Quebec it
is agressively priced at $1100 CND with 4megs on it. Needless to say
the board is selling like hot cakes here.
I have one and have been rendering hundreds of frames on it with
(about to be released) Caligari-Broadcast2.1 without a single
failure with rendering times around 4x faster than on the A3000/25 ,
which I also have.
If by "old" you mean they were there first, that may be true but if
by "old" you mean obsolete, forget it. This board is *fast* and it's
agressively priced with good customer support.
greg schaem
Sory You didn't get it.50mhz bus 50mhz internal.....
I was talking about 25mhz/50mhz 040 and said 'WHAT a
50mhz bus speed would help in a A3000'.I didn't say
double the internal clock!
From your word a 50mhz/50mhz 040 is 0% faster than
curent 25mhz 040.So why people are so hunappy and
mad when say the 50mhz ads are hypes! 50mhz bus wont
help... You would need 20ns RAM! (I asume so since
40ns ram give 0 wait state to 25mhz 040).
And Wrong the cache HAS NOTHING to do with it.Unless
the bus clock is used by the sequencencer.
So a 33mhz (Shiping only in the past months) should give
you around 30% when the Bus controller is inactive.
The actual relation to refresh the memory was: was
has the cache to do with it... In realtion with the
BUS speed in a A3000.
To be very simple should motorola make a 040 with a
50mhz bus speed? And if they do why amiga owner should
pay extra for it?
I already see the type of asnwer so: No I didn't say Moto
should not make 100mhz 040 with internal clock.
Also for information, I do timing to optimize code on
68020+ routine, taking into acount prefect, instruction
overlap, concurancy, bus usage, cache size, etc...
The only answer I was looking for was if th 040 had a 50mhz
internal clock or the 50mhz was really only hipe.
Stephan.
Ronald van Eijck
>Assume you could get a 50/100MHz '040 (you can't, the fastest available is
>the 33/66MHz part). It takes just as long as the 25MHz part to read or write
>external RAM. Anything in cache, and any internal operation as well, goes
>twice as fast. If you consider that this hit rate on an '040 cache is way
>up there (I don't know offhand -- 80%? 90%? 95%? Mike?), you'll find that a
>good portion of the time the faster processor would indeed go faster. Since
>it'll never go slower, you'll get an improvement. Sure, it may not be 100%,
>but 70% is nice.
>--
>Dave Haynie Commodore-Amiga (Amiga 3000) "The Crew That Never Rests"
> {uunet|pyramid|rutgers}!cbmvax!daveh BIX: hazy
> "I used to be disgusted, now I try to be amused" - Elvis Costello
Looking at this data wouldn't it be wiser for motorola to make a 25/100 Mhz
68040 since doubling the speed of the memory bus is not going to speedup
operations wouldn't it be best to keep it at 25 Mhz, This way the high
frequency signals would be inside the chip and could be shielded and it
would remove the problems of designing 50 Mhz logic on the mainboard.
Ofcourse as soon as Faster ram is available the 50/100 would be a better
option but until then the only real benefit of a 50/100 over a 25/100 would
be that you could improve speed using a fast large external cache.
C'ya,
--
+-------------------------------------------------------------------------+
| Ronald van Eijck R&R Software |
| |
| You move slowly, but surely, to success. (my fortune cookie) |
+-------------------------------------------------------------------------+
Tommy Petersson
- Dan_...@cave.UUCP (Dan Green) writes:
- >[lines deleted]
- >moment, Fusion 40 is an old one, but it's still an option (but you can
- >forget about the 50MHz version :-)).
-
- >wr.
- > _________________________
- > / Dan Green / Sprintsoft /
- >/________________________/
- >Dan_Green%ca...@sweden.eu.net
-
- >-- Via DLG Pro v0.985b
-
- The Fusion40 is also now 28mhz with upto 32megs onboard. I believe it's
- the least expensive board as far as the US goes and here in Quebec it
- is agressively priced at $1100 CND with 4megs on it. Needless to say
- the board is selling like hot cakes here.
How much is CND 1100? Is it about USD 1000? This seems VERY agressive,
just about the price originally mentioned for the '040 cards. Is it
also out in a A3000 version?
--
=============================================================================
Tommy Petersson to...@enea.se Enea Data AB, Sweden
=============================================================================
Tommy Petersson
easily disable the board (like by pressing the mouse buttons on
boot-up)? I will be able to run Unix on the normal '030 in my A3000
until Amiga Unix hopefully will be upgraded to handle 68040.
Dave Haynie
>In article <29...@cbmvax.commodore.com> da...@cbmvax.commodore.com (Dave Haynie) writes:
>>Assume you could get a 50/100MHz '040 (you can't, the fastest available is
>>the 33/66MHz part). It takes just as long as the 25MHz part to read or write
>>external RAM.
>Looking at this data wouldn't it be wiser for motorola to make a 25/100 Mhz
>68040 since doubling the speed of the memory bus is not going to speedup
>operations wouldn't it be best to keep it at 25 Mhz,
Not necessarily. First of all, Motorola only wants to make so many '040 varities.
If you given me a 50/100MHz part, I don't have to have much of the external bus
run at 50MHz; memory and everything else could easily use a synchronous 25MHz
clock, keeping things cheap and quiet. The 25/100MHz part might give the same
advantages in this system, but if I could get faster DRAM (RAMBUS, anyone?) or a
good external cache design, this wouldn't help me, while the former part would.
The main advantage to a doubling scheme like the '040 (well, not really doubling;
you have to provide both clocks on the motherboard, though PCLK need only go to
the '040) is that internal operations happen so much faster than external
operations, you can support the increased speed in the same level of chip
technology needed to get the external bus going half that speed. This wasn't
practical in the past, since the CPU clock tended to run everything very
tightly coupled in a system. Modern CPUs like the '040 tend to have a bus
interface unit which is responsible for I/O between the bus and the caches,
and various somewhat-decoupled functional units which expect to fetch from cache.
This way the high
>frequency signals would be inside the chip and could be shielded and it
>would remove the problems of designing 50 Mhz logic on the mainboard.
>
>Ofcourse as soon as Faster ram is available the 50/100 would be a better
>option but until then the only real benefit of a 50/100 over a 25/100 would
>be that you could improve speed using a fast large external cache.
>
>C'ya,
>
>--
> +-------------------------------------------------------------------------+
> | Ronald van Eijck R&R Software |
> | |
> | You move slowly, but surely, to success. (my fortune cookie) |
> +-------------------------------------------------------------------------+
greg schaem
Nope, since the begining it was a BUS VS SEQUENCER timing
(has I remenber writting about).Anyway I know that having faster
speed make the chip go faster, having a faster buss with curent
memory is another question.It's not my area, and could think that
that a 50mhz bus speed could increase speed a bit because of
Async access (Or am I off the track?)
And curently the A3000 does not offer External cache, etheir has
the 040 boards? But it would be nice to have the 128K Motorola cahce
inplementation. (When the 50mhz/100mhz 040 arive:-)
Stephan.
Rostyk Lewyckyj
>but if I could get faster DRAM (RAMBUS, anyone?) or a
>good external cache design,
Dave, I am sure that you know about memory interleave as a standard
solution for building faster memory systems out of slow memory chips.
So why do we need to wait for faster DRAM or for a "good cache design"
It seems to me that caches are needed when the speed difference between
the available parts and the desired access speed is too many factors
different to match via memory interleave. But in this case we are talking
about a factor of 2 or at most 4.
Please explain why memory interleave is not an acceptable solution.
Use 256Kx4 80ns parts and 4 way interleave. Then the system would
still be expandable in 4M steps. Sure it would cost some real estate
to use 256Kx4 vs. 1Mx4 chips, and some glue logic, and some thinking
design time.
But we know you could do it.
Stephen Menzies
The RCS Fusion40 runs on the A2000 only. It has a hardware switch to go
back and forth from 68000 to 040. Never tried it though.
-stephen.
Stephen Menzies
>men...@CAM.ORG (Stephen Menzies) writes:
>-
>- The Fusion40 is also now 28mhz with upto 32megs onboard. I believe it's
>- the least expensive board as far as the US goes and here in Quebec it
>- is agressively priced at $1100 CND with 4megs on it. Needless to say
>- the board is selling like hot cakes here.
>How much is CND 1100? Is it about USD 1000? This seems VERY agressive,
>just about the price originally mentioned for the '040 cards. Is it
>also out in a A3000 version?
>--
>=============================================================================
> Tommy Petersson to...@enea.se Enea Data AB, Sweden
>=============================================================================
This CND price is a limited time offer to Quebec club members and
developers. From what I understand the special is supposed to be offered in
the US (club members/dev) very soon. The prices will be very agressive
although I'm not so sure they will quite this good. When the "special" is
offered I will announce it and give full details.
And yes the CND is worth about 84c US.
No, there's no A3000 version.
Dave Haynie
>In article <29...@cbmvax.commodore.com>, da...@cbmvax.commodore.com (Dave Haynie) writes:
>>but if I could get faster DRAM (RAMBUS, anyone?) or a
>>good external cache design,
>Dave, I am sure that you know about memory interleave as a standard
>solution for building faster memory systems out of slow memory chips.
Of course, it's a standard trick. Last time I tried it was on the early
A2630 prototypes. It turns out (and it's still basically true) that you
can't count on using the DRAM's own output buffer on/off times for anything
that really needs the speed of interleave. So you need buffers...
>So why do we need to wait for faster DRAM or for a "good cache design"
>It seems to me that caches are needed when the speed difference between
>the available parts and the desired access speed is too many factors
>different to match via memory interleave. But in this case we are talking
>about a factor of 2 or at most 4.
Well, you can go either way. In many cases, we're constrained to the specs
the management want. It wouldn't be any technical problem to build a four
way interleaved memory system. But it has its share of problems. You four
times the buffers (16 of them, or 8 if you use the more expensive WideBus(tm)
buffers; that's as much as $100 extra cost at retail if you go for the more
sophisticated latching buffers we generally find useful) and four times the
memory and buffer control lines. The grain of your memory goes up by four,
so instead of supporting 1MB increments, you may be required to support 4MB
increments, depending on the flexibility in your controller (eg, can it
support non-interleaved operation). Then again, there's not all that much
use for a four-way interleaved memory system anyway. Interleaved systems are
good for supporting burst cycles, which we do via static column modes anyway.
A two-way interleaved system with the same kind of static column functions we
use could give you 1 clock burst rather than 2 clock burst, that's probably
a 10%-20% speedup. Two-way interleave also lets you hide precharge time if
your accesses line up right, which can be a substantial improvement, as things
go closer to access time than cycle time. Four-way interleave can also be
used for burst support as I did in the A2630; it's often easier than building
a fast-page or SCRAM control circuit, especially if you design in PALs. Beyond
that, interleave has some applications in muliprocessing systems and other
places you need bandwidth improvements beyond what the CPU needs, but you
don't get all that much more increase for the costs going up in interleave
with a single CPU.
The appeal of a memory solution rather than a system solution is that you may
see much the same performance upgrade for not much extra cost, or maybe even
less. These caching DRAM are making a big noise because they can easily fit
in existing architectures and they offer a good speedup per dollar. The new
synchronous DRAM should have a similar effect; they could even lower the cost
of a bursting/interleaved system, since their output buffers should be faster
than those of traditional DRAM and they latch data, so at the very least they
eliminate the more exotic buffers from your DRAM system (74F646 and 74F543
are our favorites here). And of course, any technique you find useful with
common DRAM are useful with clever new ones. The appeal of faster memory, of
course, is that it doesn't have to raise the cost of your system. You can
ship with common DRAM, or small amounts, like most systems houses do, and let
the user's pay whatever upgrade parts they can afford. You don't generally
factor next year's upgrade into the price you pay today, so support of faster
drop-in replacements doesn't make you look expensive w.r.t. the competition.
People don't pay much attention to things they don't understand, unless they're
easy. CPU clock speed is a big one, and simple numbers like cache size and
DRAM rating are things semi-educated buyers look at, but there's often little
way to attract $100's worth of attention to a really fast DRAM system. Which
is why so many vendors build mediocre DRAM systems -- after all, most benchmarks
fit in cache anyway :-) (incidently, I certainly don't think this way for my
systems, but I understand the offtimes marketing imposed reasoning for it in
the computer marketplace).