Secret why Pentium beats K6 on Quake
Victor Major
Victor Major <stm...@gsb.usyd.edu.au> write:
> Darrin wrote:
> On Wed, 10 Dec 1997 10:07:19 GMT, hi...@uoguelph.ca (Anthony Hill)
> wrote:
>
> >
> > Quake uses the FPU for nearly all the graphics stuff it does.
> >In fact, what it tends to do is bombard the FPU with instruction after
> >instruction pretty consistantly for some points.
This is complete crap. Qauke uses the FPU only to add two long strings in
the FPU registers since they (iD) could not be bothered to rewrite that
part of code for pentiums which can add long strings in the integer
registers. There is absolutely NO FPU processing bar addition that is
performed by Quake in the FPU!!! The reason why AMD is slower is because
it cannot move data in and out of those registrs as fast as the pentium
FPU.
V.
-------------------==== Posted via Deja News ====-----------------------
http://www.dejanews.com/ Search, Read, Post to Usenet
Micheal Cranford
Someone (correctly) wrote:
"Quake uses the FPU for nearly all the graphics stuff it does.
In fact, what it tends to do is bombard the FPU with instruction
after instruction pretty consistantly [sic] for some points."
Victor Major <stm...@gsb.usyd.edu.au> responded:
"This is complete crap. Qauke [sic] uses the FPU only to add two
long strings [sic] in the FPU registers since they (iD) could not
be bothered to rewrite that part of code for pentiums [sic] which
can add long strings [sic] in the integer registers. There is
absolutely NO FPU processing bar addition that is performed by
Quake in the FPU!!!"
The above paragraph contains so many errors that I assume it is a troll. I
am writing a correction for those readers who are not familiar with the topic
being discussed.
First, Quake (I assume that was what was meant by "Qauke") uses Floating-
Point Addition (FADD), Multiplication (FMUL) and Division (FDIV), in addition
to loads and stores. I know this because I have the source code since it was
sent to me by the Michael Abrash (the author of the graphics routines at iD).
Others can easily verify that these instructions are used by disassembling the
Quake executable. By the way, there are 50% more FMUL instructions than FADD
instructions in the Quake graphics code so FADD isn't even in the majority.
Second, "add two long strings in the FPU registers" does not make any sense.
How does one "add long strings" (long sequences of ASCII characters) in the FPU
registers?
Third, how can anyone "rewrite that part of code for pentiums which can add
long strings in the integer registers" when the code was originally written for
the Pentium (R) and all Pentiums have identical integer register capability? In
other words, no such rewrite is even possible.
Fourth, "There is absolutely NO FPU processing bar addition that is performed
by Quake in the FPU!!!" was partially addressed in my first point. Since there
are 22 FADD instructions and 39 non-FADD Floating-Point math instructions (I do
not count loads and stores as math instructions) it is also interesting to note
that *most* of the FPU math instructions in Quake are *not* FADD.
"The reason why AMD is slower is because it cannot move data in and
out of those registrs [sic] as fast as the pentium FPU."
That makes zero out of five correct claims. The Pentium has a pipelined FPU
while the K6 does not (assuming that was what was meant by "AMD"). This gives
the Pentium a higher possible FPU throughput.
--
******************************************************************************
Micheal Cranford Resident Skeptic Intel P6 Architecture Labs
Chris Lomont
Victor Major wrote in message <882074588...@dejanews.com>...
>Victor Major <stm...@gsb.usyd.edu.au> write:
>
>> Darrin wrote:
>> On Wed, 10 Dec 1997 10:07:19 GMT, hi...@uoguelph.ca (Anthony Hill)
>> wrote:
>>
>> >
>> > Quake uses the FPU for nearly all the graphics stuff it
does.
>> >In fact, what it tends to do is bombard the FPU with instruction
after
>> >instruction pretty consistantly for some points.
>
>This is complete crap. Qauke uses the FPU only to add two long
strings in
>the FPU registers since they (iD) could not be bothered to rewrite
that
>part of code for pentiums which can add long strings in the integer
>registers. There is absolutely NO FPU processing bar addition that is
>performed by Quake in the FPU!!! The reason why AMD is slower is
because
>it cannot move data in and out of those registrs as fast as the
pentium
>FPU.
This is completely wrong. Abrash, one of the programmers of Quake,
gave a talk at the Computer Game Developers Conference detailing the
use of the FPU in Quake. He wrote articles in Doctor Dobbs Journal
showing FPU optimizations he used for Quake for matrix multiplies and
dot products. He detailed how all the geometry is done quicker with
the FPU than any other method he could find. And he also wrote that he
is glad the days of fixed point are over.
Where did you get the idea they use it for "long string" addition,
whatever that would be used for?
AMD is slower because it cannot do FXCH for free in the manner the
Pentium does. Thus Quake, and any game heavily optimized for Intel
FPU, runs like crap on AMD and Cyrix. Newer generations of chips may
address this.
Chris Lomont
Russ Williams
Chris Lomont wrote in message <66vrhq$a...@mozo.cc.purdue.edu>...
[...]
>AMD is slower because it cannot do FXCH for free in the manner the
>Pentium does. Thus Quake, and any game heavily optimized for Intel
>FPU, runs like crap on AMD and Cyrix. Newer generations of chips may
>address this.
I wish everyone would stop lumping AMD in with Cyrix. I use a K6-166
and it runs Quake as well as a P166. AMD chips may not be quite as
fast as Intel's, but it's *close*.
From the specs on AMD's site, it seems that (on a K6) fmul/fadd
instructions take 2/2 clocks instead of 3/1 on a Pentium. That means
that although the instruction completes quicker, there is no opportunity
to call fxch. The 6x86, OTOH, takes (IIRC) 10 clocks for an fmul. That
with the overrating of the clock speed (133MHz => PR166) makes a
Cyrix 6x86 utterly shite for Quake.
---
Russ
Patrick
On 14 Dec 1997, Micheal Cranford wrote:
[...]
> That makes zero out of five correct claims. The Pentium has a pipelined FPU
> while the K6 does not (assuming that was what was meant by "AMD"). This gives
> the Pentium a higher possible FPU throughput.
[...]
This just about seems to sum up the largest problem with the K6 as it now
stands. It would be this non-pipelined fpu that also makes the K6 slower
on rendering raytraced images (?).
I hope that the K6 3D and K6+3D will correct this deficiency. The fpu
and only the fpu is what has held me back from going K6 thus far.
patrick
sumatose
Victor Major wrote:
> This is complete crap. Qauke uses the FPU only to add two long strings in
> the FPU registers since they (iD) could not be bothered to rewrite that
> part of code for pentiums which can add long strings in the integer
> registers. There is absolutely NO FPU processing bar addition that is
> performed by Quake in the FPU!!! The reason why AMD is slower is because
> it cannot move data in and out of those registrs as fast as the pentium
> FPU.
All the 3D math in Quake is done in floating point. That's just what the
facts are. ID software has said from the start that floating point
math is faster than fixed-point on a pentium.
<SARCASM> That AMD succeeded to make a floating-point unit slower than
Intel - who was re-knowned to make slow ones - is quite an achievement.
I hope you're enjoying the 30$ you saved by buying a non-intel cpu! ;-) </SARCASM>
Jim
On Sun, 14 Dec 1997 16:39:51 -0500, sumatose <suma...@NOSPAM.usa.net>
wrote:
><SARCASM> That AMD succeeded to make a floating-point unit slower than
>Intel - who was re-knowned to make slow ones - is quite an achievement.
>I hope you're enjoying the 30$ you saved by buying a non-intel cpu! ;-) </SARCASM>
Gee, I hope you're enjoying you more expensive Intel chip, that gets 2
more fps in Quake, but is slower running everything else. <G>
Stephan Pelletier
shawn
Great! I think I will run out and buy a over priced CPU by a greedy company
(Intel) so that I can get 2 more FPS out of a piece of shit game (quake..or
quack). Hell even GLQuake looks like ass. If games are that damn important
to everyone why the hell did you buy a computer anyway, shit get a
nintendo64 and monitor frame rates all day. Sorry about the ranting and
raving but these dipshit threads buy people with the tech knowledge of a 5
YO are getting on my nerves. If you don't know the answer don't just f'n
guess.....Shawn
Victor Major wrote in message <882074588...@dejanews.com>...
>Victor Major <stm...@gsb.usyd.edu.au> write:
>
>> Darrin wrote:
>> On Wed, 10 Dec 1997 10:07:19 GMT, hi...@uoguelph.ca (Anthony Hill)
>> wrote:
>>
>> >
>> > Quake uses the FPU for nearly all the graphics stuff it does.
>> >In fact, what it tends to do is bombard the FPU with instruction after
>> >instruction pretty consistantly for some points.
>
>the FPU registers since they (iD) could not be bothered to rewrite that
>part of code for pentiums which can add long strings in the integer
>registers. There is absolutely NO FPU processing bar addition that is
>performed by Quake in the FPU!!! The reason why AMD is slower is because
>it cannot move data in and out of those registrs as fast as the pentium
>FPU.
>
Blue Nova
>Victor Major <stm...@gsb.usyd.edu.au> write:
>
>> Darrin wrote:
>> On Wed, 10 Dec 1997 10:07:19 GMT, hi...@uoguelph.ca (Anthony Hill)
>> wrote:
>>
>> >
>> > Quake uses the FPU for nearly all the graphics stuff it does.
>> >In fact, what it tends to do is bombard the FPU with instruction after
>> >instruction pretty consistantly for some points.
>
>This is complete crap. Qauke uses the FPU only to add two long strings in
>the FPU registers since they (iD) could not be bothered to rewrite that
>part of code for pentiums which can add long strings in the integer
>registers. There is absolutely NO FPU processing bar addition that is
>performed by Quake in the FPU!!! The reason why AMD is slower is because
>it cannot move data in and out of those registrs as fast as the pentium
>FPU.
>
> V.
>
>-------------------==== Posted via Deja News ====-----------------------
> http://www.dejanews.com/ Search, Read, Post to Usenet
On a similar, but unrelated note, I remember in an old issue of PC
Gamer, before Quake was released, it said, and I quote:
"Because Quake will be using the FPU unit, or math coprocessor, it
will be about to run well on even 486 CPU's."
What a joke!
T.W. Seddon
sumatose (suma...@NOSPAM.usa.net) wrote:
> All the 3D math in Quake is done in floating point. That's just what the
> facts are. ID software has said from the start that floating point
> math is faster than fixed-point on a pentium.
> <SARCASM> That AMD succeeded to make a floating-point unit slower than
> Intel - who was re-knowned to make slow ones - is quite an achievement.
> I hope you're enjoying the 30$ you saved by buying a non-intel cpu! ;-) </SARCASM>
Plus ca change...
People argued this shite to death when it was first noticed that Quake was
slow on the Cyrix 6x86. The argument went round and round and round and round
and round and round and round and round and round and round and round and
round in circles, in a fashion as repetitive as this sentence.
Now go outside! Get some fresh air! Meet people! Quake is just a game, for
God's sake.
--Tom
Jason A. Petrasko
I will be waiting to see Cyrix's "Cayenne" scheduled from mid-1998
release. They are supposed to improve the FPU. Hopefully, they will
improve it better than they did with their new MMX engine in the M2.
Which, by the way, got the worst MMX test scores of all MMX CPUS...
Skeletonmaster
not all true, its "semi-decent" for quake, i run on 320x200 anyway and get 30-40
fps, i dont like the look of the higher res. anyway (shrunken ammo bar, etc. ya
even 640x480)My Cyrix P-200+ is fine, and it runs all my other stuff real fast.
(¯`·.¸¸.·´¯`·->«««· Skeletonmaster ·»»»<-·´¯`·.¸¸.·´¯)
|§¤*~*¤§|§¤*~ - - skeleto...@cyberdude.com - - ~*¤§|§¤*~*¤§|
©®©®©®©®©® http://members.tripod.com/~skeletonmaster ©®©®©®©®©®
(¯`·.¸¸.·´¯`·->«««· My ICQ UIN is: 2076577 ·»»»<-·´¯`·.¸¸.·´¯)
Russ Williams wrote in message <670kfc$j...@usenet42.supernews.com>...
hades
No.
QUAKE isn't just a game.
David
On 15 Dec 1997 18:15:17 GMT, "T.W. Seddon" <T.W.S...@ncl.ac.uk>
wrote:
>People argued this shite to death when it was first noticed that Quake was
>slow on the Cyrix 6x86. The argument went round and round and round and round
>and round and round and round and round and round and round and round and
>round in circles, in a fashion as repetitive as this sentence.
>
>Now go outside! Get some fresh air! Meet people! Quake is just a game, for
>God's sake.
>
>--Tom
GLquake is fairly awesome on a 6x86 (PR200+) and 3dfx, so there is
hope for cyrix owners.
James Boswell
>Which, by the way, got the worst MMX test scores of all MMX CPUS
Which IIRC was about 2% _SLOWER_ than the cyrix for FP execution!!
-_-_
James Boswell|James -at- hillbanks.demon.co.uk (you know the drill,
replace -at- with @ etc.)
*They make you suffer, they make you twist and scream,*
* they are the enemy, and the war has just begun* - Demon's Within
*well, I guess trying to explain the concept of intelligence to you is like*
*trying to explain the concept of a chord to greenday...* - Seroussi
Anthony Hill
On Mon, 15 Dec 1997 19:52:53 -0500, "Jason A. Petrasko"
<kag...@ameritech.net> wrote:
>> I hope that the K6 3D and K6+3D will correct this deficiency. The fpu
>> and only the fpu is what has held me back from going K6 thus far.
>>
>
>release. They are supposed to improve the FPU. Hopefully, they will
>improve it better than they did with their new MMX engine in the M2.
I wouldn't really go that far at this time. Currently there
are NO MMX benchmark suits availible. The closest thing is Intel's
Multimedia Benchmark. While this benchmark is often refered to as an
MMX benchmark, it actually uses the FPU of the processor more then the
MMX insturction set. In some real-world applications Cyrix turned in
some MMX scores which were better then PIIs, while in other
applications it didn't far so well. As of now it's pretty much
impossible to judge who has the best or worst MMX unit of the bunch.
Anthony Hill | Sig files? SIG FILES?! What the
hi...@uoguelph.ca | hell do I need a sig file for?!
Rob Rodgers
"James Boswell" <beez...@department-of-torture.hell.gov> wrote:
>>Which, by the way, got the worst MMX test scores of all MMX CPUS
>Not counting the IDT WINCHIP C6
>
>Which IIRC was about 2% _SLOWER_ than the cyrix for FP execution!!
That's tough. Did the chip start sliding backwards in spacetime or
something?
----
http://www.wam.umd.edu/~rsrodger
Odd news in computer history: (2/11/97) Sun announces
bizarre plan to "rescue" and "revive" old, slow 486 systems
by having them run Java.
Tim Smith
Victor Major <victor...@hotmail.com> wrote:
[wrong stuff deleted]
Mr. Major has no idea how Quake works. Ignore him, and instead read the
articles by Michael Abrash in Dr. Dobbs where Abrash talks about various
optimizations in Quake, or read assorted past .plan updates from Carmack
where he discusses Quake and floating point, or disassemble Quake.
(Or get a motherboard with an LM78, and watch the temperature go up by 5
degrees celsius after an afternoon of Quake :-) ).
--Tim Smith
|fREoN|
It doesnt matter anyway because the Pentium doesnt beat the K6 in Quake.
Well it does, but only by about 1fps or less. It was tested on one system,
same config, just swapped the CPU's.
Tim Smith wrote in message <67dab5$5fa$1...@halcyon.com>...
Daniel
If we did not have K6 today, you would pay $$$10,000 for P166MMX and P2 would not be
even out yet. Sooooo stop bitching people !!!!!!!!!
Shawn
Daniel wrote:
: If we did not have K6 today, you would pay $$$10,000 for P166MMX and P2 would not be
: even out yet. Sooooo stop bitching people !!!!!!!!!
If we had windows 99 written & sold by xxx then microsoft would be in
BIG trouble. Competetion drives the prices down & prevents gouging.
bri...@proaxis.com
Anthony Hill wrote:
This is not eniterly true. Norton Utilities for Windows 95 v2.0
contains the Norton Media Benchmark which is MMx enabled. Unfortunately,
it also requires a CD-ROM drive and sound card installed on the system.
Brian Tottleben
bri...@proaxis.com
Yit Loong Lai
It's nothing to do with CPU. Competition among PC manufacturers drive down
the price. You pay $10,000 in those days when IBM was the only PC
manufacturer, now you have Compaq, DELL, Micro and you can even build your
own if you wish!
Shawn wrote in message <67mh43$4...@news.southeast.net>...
Doug
"Yit Loong Lai" <lyl...@pc.jaring.my> wrote:
>It's nothing to do with CPU. Competition among PC manufacturers drive down
>the price. You pay $10,000 in those days when IBM was the only PC
>manufacturer, now you have Compaq, DELL, Micro and you can even build your
>own if you wish!
>
Really?? Nothing?? Nothing at all?? The fact that AMD and Cyrix
make x86-compatible processors doesn't affect the price the the PC at
all??
Chris Quirke
>>Daniel wrote:
>>: If we did not have K6 today, you would pay $$$10,000 for P166MMX and P2
>>:would not be even out yet. Sooooo stop bitching people !!!!!!!!!
>>Competetion drives the prices down & prevents gouging.
>It's nothing to do with CPU. Competition among PC manufacturers drive down
>the price. You pay $10,000 in those days when IBM was the only PC
>manufacturer, now you have Compaq, DELL, Micro and you can even build your
>own if you wish!
When Intel sells chips that cost half as much to make at double the
cost, and incidentally more than a complete entry-level system, it has
*everything* to do with the CPU.
The purpose of Intel's marketing is not only to lure sales from AMD
and Cyrix, it is to lure value-add expenditure from SVGA, hard drive
and other component sectors for a larger slice of the total-system
piechart. Something is seriously wrong when the "new generation"
heralds 33% of system price for the CPU, L2 cache removed from
motherboard to CPU, and *weaker* peripherals such as "soft" modems,
sound cards and video that increase the need for CPU performance.
After all, "A computer is only as good as it's processor", right?
>----------- ------ ----- ---- --- -- - - - - -
You know you need a holiday when...
...25 December is just another VAT deadline
Tillmann Steinbrecher
Hi,
Chris Quirke wrote:
> When Intel sells chips that cost half as much to make at double the
> cost, and incidentally more than a complete entry-level system, it has
> *everything* to do with the CPU.
>
> The purpose of Intel's marketing is not only to lure sales from AMD
> and Cyrix, it is to lure value-add expenditure from SVGA, hard drive
> and other component sectors for a larger slice of the total-system
> piechart. Something is seriously wrong when the "new generation"
> heralds 33% of system price for the CPU, L2 cache removed from
> motherboard to CPU, and *weaker* peripherals such as "soft" modems,
> sound cards and video that increase the need for CPU performance.
>
> After all, "A computer is only as good as it's processor", right?
Excellent post. Why should you let a $10 DSP do something your
$500 CPU can do as well?
And, it's not only intel who's acting that way - have a
look at the Creative AWE64 - it's a 32-channel soundcard,
and the other 32 channels are software-emulated via
the CPU - if it's fast enough.
regards,
Tillmann
--
_____ ___ _____
|_ _| / __|_ _| Tillmann Steinbrecher Black holes are
| | _ \__ \ | | _ mailto://t...@gmx.de where God
|_|(_)|___/ |_|(_) http://till.home.ml.org divided by zero