Article: PhysX87: Software Deficiency
By: David Kanter (dkanter.delete@this.realworldtech.com), July 7, 2010 2:55 pm
Room: Moderated Discussions
Joel Hruska (joel.hruska@gmail.com) on 7/7/10 wrote:
---------------------------
>David,
>
>I'm wondering if you had a chance to read over the whitepaper Intel recently published
>comparing CPU and GPU performance:
>
>http://portal.acm.org/citation.cfm?id=1816021&coll=GUIDE&dl=GUIDE&CFID=94608761&CFTOKEN=50783980&ret=1#Fulltext
>
>Their work seems to agree with some of your findings, but >it's in relation to CUDA, not PhysX.
>
>My first question is whether or not you see your >investigation reinforcing Intel's
>own paper, or if they're two different topics and >computing areas.
I think they are related but different. I focused on why we see a performance gap between CPU and GPU in a specific case (e.g. PhysX).
I haven't read Intel's paper yet, but I believe their central point is that many of the CPU vs. GPU comparisons are not apples to apples (e.g. one is more optimized than the other, as is the case with PhysX). Moreover, I think they are arguing that if you do apples to apples comparisons, the performance gaps are not nearly as big.
I think my work is roughly aligned with some of what they are doing and seems congruent. Certainly I wouldn't consider PhysX an apples to apples comparison. But I don't have the ability to really examine what the performance gap for PhysX really is. Perhaps I can get to that sooner or later.
>My second question is this: In this article, you discuss x87 vs. SSE/SSE2, but
>don't mention SSE3, SSE4.1, or SSE4.2. It was always my >understanding that SSE2
>was the most important of Intel's SIMDs, but I never knew >if that was because it
>came out at the same time as the P4, and had to be used >for any kind of decent performance.
SSE1 was mostly for 4x32b (single precision), which was great for some things. But if you wanted 64 bits (double precision), you were stuck. And x87 was 80 bit precision.
So SSE1 couldn't totally replace x87. However, SSE2 makes x87 totally obsolete since it does 2x64b. That's why it's a big deal.
SSE3 and onwards were really more of refinements that added some missing instructions to SSE1 and SSE2 and some special purpose ones as well (e.g. string instructions in SSE4.2). But SSE1+SSE2 are really the key.
>How important are SSE3/4.1/4.2, and will AVX actually >improve SSE2 performance
>by handling instructions more efficiently?
AVX extends the vector length from 128 bits to 256-bits. So in theory, you could do 8x32b or 4x64b operations per cycle; basically double what is possible with SSE. In practice, the improvement probably isn't 2X and is probably closer to 1.5X...but that's due to the hardware that will run AVX, and not really a function of the instructions.
Comparing x87 to AVX, the improvement is probably something like 2-4X.
>Thanks much.
Sure thing. You should try and link up the original piece at HH : )
David
---------------------------
>David,
>
>I'm wondering if you had a chance to read over the whitepaper Intel recently published
>comparing CPU and GPU performance:
>
>http://portal.acm.org/citation.cfm?id=1816021&coll=GUIDE&dl=GUIDE&CFID=94608761&CFTOKEN=50783980&ret=1#Fulltext
>
>Their work seems to agree with some of your findings, but >it's in relation to CUDA, not PhysX.
>
>My first question is whether or not you see your >investigation reinforcing Intel's
>own paper, or if they're two different topics and >computing areas.
I think they are related but different. I focused on why we see a performance gap between CPU and GPU in a specific case (e.g. PhysX).
I haven't read Intel's paper yet, but I believe their central point is that many of the CPU vs. GPU comparisons are not apples to apples (e.g. one is more optimized than the other, as is the case with PhysX). Moreover, I think they are arguing that if you do apples to apples comparisons, the performance gaps are not nearly as big.
I think my work is roughly aligned with some of what they are doing and seems congruent. Certainly I wouldn't consider PhysX an apples to apples comparison. But I don't have the ability to really examine what the performance gap for PhysX really is. Perhaps I can get to that sooner or later.
>My second question is this: In this article, you discuss x87 vs. SSE/SSE2, but
>don't mention SSE3, SSE4.1, or SSE4.2. It was always my >understanding that SSE2
>was the most important of Intel's SIMDs, but I never knew >if that was because it
>came out at the same time as the P4, and had to be used >for any kind of decent performance.
SSE1 was mostly for 4x32b (single precision), which was great for some things. But if you wanted 64 bits (double precision), you were stuck. And x87 was 80 bit precision.
So SSE1 couldn't totally replace x87. However, SSE2 makes x87 totally obsolete since it does 2x64b. That's why it's a big deal.
SSE3 and onwards were really more of refinements that added some missing instructions to SSE1 and SSE2 and some special purpose ones as well (e.g. string instructions in SSE4.2). But SSE1+SSE2 are really the key.
>How important are SSE3/4.1/4.2, and will AVX actually >improve SSE2 performance
>by handling instructions more efficiently?
AVX extends the vector length from 128 bits to 256-bits. So in theory, you could do 8x32b or 4x64b operations per cycle; basically double what is possible with SSE. In practice, the improvement probably isn't 2X and is probably closer to 1.5X...but that's due to the hardware that will run AVX, and not really a function of the instructions.
Comparing x87 to AVX, the improvement is probably something like 2-4X.
>Thanks much.
Sure thing. You should try and link up the original piece at HH : )
David
Topic | Posted By | Date |
---|---|---|
A bit off base | John Mann | 2010/07/07 07:04 AM |
A bit off base | David Kanter | 2010/07/07 11:28 AM |
SSE vs x87 | Joel Hruska | 2010/07/07 12:53 PM |
SSE vs x87 | Michael S | 2010/07/07 01:07 PM |
SSE vs x87 | hobold | 2010/07/08 05:12 AM |
SSE vs x87 | David Kanter | 2010/07/07 02:55 PM |
SSE vs x87 | Andi Kleen | 2010/07/08 02:43 AM |
80 bit FP | Ricardo B | 2010/07/08 07:35 AM |
80 bit FP | David Kanter | 2010/07/08 11:14 AM |
80 bit FP | Kevin G | 2010/07/08 02:12 PM |
80 bit FP | Ian Ollmann | 2010/07/19 12:49 AM |
80 bit FP | David Kanter | 2010/07/19 11:33 AM |
80 bit FP | Anil Maliyekkel | 2010/07/19 04:49 PM |
80 bit FP | rwessel | 2010/07/19 05:41 PM |
80 bit FP | Matt Waldhauer | 2010/07/21 11:11 AM |
80 bit FP | Emil Briggs | 2010/07/22 09:06 AM |
A bit off base | John Mann | 2010/07/08 11:06 AM |
A bit off base | David Kanter | 2010/07/08 11:27 AM |
A bit off base | Ian Ameline | 2010/07/09 10:10 AM |
A bit off base | Michael S | 2010/07/10 02:13 PM |
A bit off base | Ian Ameline | 2010/07/11 07:51 AM |
A bit off base | David Kanter | 2010/07/07 09:46 PM |
A bit off base | Anon | 2010/07/08 12:47 AM |
A bit off base | anon | 2010/07/08 02:15 AM |
A bit off base | Gabriele Svelto | 2010/07/08 04:11 AM |
Physics engine history | Peter Clare | 2010/07/08 04:49 AM |
Physics engine history | Null Pointer Exception | 2010/07/08 06:07 AM |
Physics engine history | Ralf | 2010/07/08 03:09 PM |
Physics engine history | David Kanter | 2010/07/08 04:16 PM |
Physics engine history | sJ | 2010/07/08 11:36 PM |
Physics engine history | Gabriele Svelto | 2010/07/09 12:59 AM |
Physics engine history | sJ | 2010/07/13 06:35 AM |
Physics engine history | David Kanter | 2010/07/09 09:25 AM |
Physics engine history | sJ | 2010/07/13 06:49 AM |
Physics engine history | fvdbergh | 2010/07/13 07:27 AM |
A bit off base | John Mann | 2010/07/08 11:11 AM |
A bit off base | David Kanter | 2010/07/08 11:31 AM |
150 GFLOP/s measured? | anon | 2010/07/08 07:10 PM |
150 GFLOP/s measured? | David Kanter | 2010/07/08 07:53 PM |
150 GFLOP/s measured? | Aaron Spink | 2010/07/08 09:05 PM |
150 GFLOP/s measured? | anon | 2010/07/08 09:31 PM |
150 GFLOP/s measured? | Aaron Spink | 2010/07/08 10:43 PM |
150 GFLOP/s measured? | David Kanter | 2010/07/08 11:27 PM |
150 GFLOP/s measured? | Ian Ollmann | 2010/07/19 01:14 AM |
150 GFLOP/s measured? | anon | 2010/07/19 06:39 AM |
150 GFLOP/s measured? | hobold | 2010/07/19 07:26 AM |
Philosophy for achieving peak | David Kanter | 2010/07/19 11:49 AM |
150 GFLOP/s measured? | Linus Torvalds | 2010/07/19 07:36 AM |
150 GFLOP/s measured? | Richard Cownie | 2010/07/19 08:42 AM |
150 GFLOP/s measured? | Aaron Spink | 2010/07/19 08:56 AM |
150 GFLOP/s measured? | hobold | 2010/07/19 09:30 AM |
150 GFLOP/s measured? | Groo | 2010/07/19 02:31 PM |
150 GFLOP/s measured? | hobold | 2010/07/19 04:17 PM |
150 GFLOP/s measured? | Groo | 2010/07/19 06:18 PM |
150 GFLOP/s measured? | Anon | 2010/07/19 06:18 PM |
150 GFLOP/s measured? | Mark Roulo | 2010/07/19 11:47 AM |
150 GFLOP/s measured? | slacker | 2010/07/19 12:55 PM |
150 GFLOP/s measured? | Mark Roulo | 2010/07/19 01:00 PM |
150 GFLOP/s measured? | anonymous42 | 2010/07/25 12:31 PM |
150 GFLOP/s measured? | Richard Cownie | 2010/07/19 12:41 PM |
150 GFLOP/s measured? | Linus Torvalds | 2010/07/19 02:57 PM |
150 GFLOP/s measured? | Richard Cownie | 2010/07/19 04:10 PM |
150 GFLOP/s measured? | Richard Cownie | 2010/07/19 04:10 PM |
150 GFLOP/s measured? | hobold | 2010/07/19 04:25 PM |
150 GFLOP/s measured? | Linus Torvalds | 2010/07/19 04:31 PM |
150 GFLOP/s measured? | Richard Cownie | 2010/07/20 06:04 AM |
150 GFLOP/s measured? | jrl | 2010/07/20 01:18 AM |
150 GFLOP/s measured? | anonymous42 | 2010/07/25 12:00 PM |
150 GFLOP/s measured? | David Kanter | 2010/07/25 12:52 PM |
150 GFLOP/s measured? | Anon | 2010/07/19 06:15 PM |
150 GFLOP/s measured? | Linus Torvalds | 2010/07/19 07:27 PM |
150 GFLOP/s measured? | Anon | 2010/07/19 09:54 PM |
150 GFLOP/s measured? | anon | 2010/07/19 11:45 PM |
150 GFLOP/s measured? | hobold | 2010/07/19 09:14 AM |
150 GFLOP/s measured? | Linus Torvalds | 2010/07/19 11:56 AM |
150 GFLOP/s measured? | a reader | 2010/07/21 08:16 PM |
150 GFLOP/s measured? | Linus Torvalds | 2010/07/21 09:05 PM |
150 GFLOP/s measured? | anon | 2010/07/22 02:09 AM |
150 GFLOP/s measured? | a reader | 2010/07/22 07:53 PM |
150 GFLOP/s measured? | gallier2 | 2010/07/23 05:58 AM |
150 GFLOP/s measured? | a reader | 2010/07/25 08:35 AM |
150 GFLOP/s measured? | David Kanter | 2010/07/25 11:49 AM |
150 GFLOP/s measured? | a reader | 2010/07/26 07:03 PM |
150 GFLOP/s measured? | Michael S | 2010/07/28 01:38 AM |
150 GFLOP/s measured? | Gabriele Svelto | 2010/07/28 01:44 AM |
150 GFLOP/s measured? | anon | 2010/07/23 04:55 PM |
150 GFLOP/s measured? | slacker | 2010/07/24 12:48 AM |
150 GFLOP/s measured? | anon | 2010/07/24 02:36 AM |
150 GFLOP/s measured? | Vincent Diepeveen | 2010/07/27 05:37 PM |
150 GFLOP/s measured? | ? | 2010/07/27 11:42 PM |
150 GFLOP/s measured? | slacker | 2010/07/28 05:55 AM |
Intel's clock rate projections | AM | 2010/07/28 02:03 AM |
nostalgia ain't what it used to be | someone | 2010/07/28 05:38 AM |
Intel's clock rate projections | AM | 2010/07/28 10:12 PM |
Separate the OoO-ness from speculative-ness | ? | 2010/07/20 07:19 AM |
Separate the OoO-ness from speculative-ness | Mark Christiansen | 2010/07/20 02:26 PM |
Separate the OoO-ness from speculative-ness | slacker | 2010/07/20 06:04 PM |
Separate the OoO-ness from speculative-ness | Matt Sayler | 2010/07/20 06:10 PM |
Separate the OoO-ness from speculative-ness | slacker | 2010/07/20 09:37 PM |
Separate the OoO-ness from speculative-ness | ? | 2010/07/20 11:51 PM |
Separate the OoO-ness from speculative-ness | anon | 2010/07/21 02:16 AM |
Separate the OoO-ness from speculative-ness | ? | 2010/07/21 07:05 AM |
Software conventions | Paul A. Clayton | 2010/07/21 08:52 AM |
Software conventions | ? | 2010/07/22 05:43 AM |
Speculation | David Kanter | 2010/07/21 10:32 AM |
Pipelining affects the ISA | ? | 2010/07/22 10:58 PM |
Pipelining affects the ISA | ? | 2010/07/22 11:14 PM |
Pipelining affects the ISA | rwessel | 2010/07/23 12:03 AM |
Pipelining affects the ISA | ? | 2010/07/23 05:50 AM |
Pipelining affects the ISA | ? | 2010/07/23 06:10 AM |
Pipelining affects the ISA | Thiago Kurovski | 2010/07/23 02:59 PM |
Pipelining affects the ISA | anon | 2010/07/24 07:35 AM |
Pipelining affects the ISA | Thiago Kurovski | 2010/07/24 11:12 AM |
Pipelining affects the ISA | Gabriele Svelto | 2010/07/26 02:50 AM |
Pipelining affects the ISA | IlleglWpns | 2010/07/26 05:14 AM |
Pipelining affects the ISA | Michael S | 2010/07/26 03:33 PM |
Separate the OoO-ness from speculative-ness | anon | 2010/07/21 05:53 PM |
Separate the OoO-ness from speculative-ness | ? | 2010/07/22 04:15 AM |
Separate the OoO-ness from speculative-ness | anon | 2010/07/22 04:27 AM |
Separate the OoO-ness from speculative-ness | slacker | 2010/07/21 07:45 PM |
Separate the OoO-ness from speculative-ness | anon | 2010/07/22 01:57 AM |
Separate the OoO-ness from speculative-ness | ? | 2010/07/22 05:26 AM |
Separate the OoO-ness from speculative-ness | Dan Downs | 2010/07/22 08:14 AM |
Confusing and not very useful definition | David Kanter | 2010/07/22 12:41 PM |
Confusing and not very useful definition | ? | 2010/07/22 10:58 PM |
Confusing and not very useful definition | Ungo | 2010/07/24 12:06 PM |
Confusing and not very useful definition | ? | 2010/07/25 10:23 PM |
Separate the OoO-ness from speculative-ness | someone | 2010/07/20 08:02 PM |
Separate the OoO-ness from speculative-ness | Thiago Kurovski | 2010/07/21 04:13 PM |
You are just quoting SINGLE precision flops? OMG what planet do you live? | Vincent Diepeveen | 2010/07/19 10:26 AM |
The prior poster was talking about SP (NT) | David Kanter | 2010/07/19 11:34 AM |
All FFT's need double precision | Vincent Diepeveen | 2010/07/19 02:02 PM |
All FFT's need double precision | David Kanter | 2010/07/19 02:09 PM |
All FFT's need double precision | Vincent Diepeveen | 2010/07/19 04:06 PM |
All FFT's need double precision - not | Michael S | 2010/07/20 01:16 AM |
All FFT's need double precision - not | Ungo | 2010/07/21 12:04 AM |
All FFT's need double precision - not | Michael S | 2010/07/21 02:35 PM |
All FFT's need double precision - not | EduardoS | 2010/07/21 02:52 PM |
All FFT's need double precision - not | Anon | 2010/07/21 05:23 PM |
All FFT's need double precision - not | Ricardo B | 2010/07/26 07:46 AM |
I'm on a boat! | anon | 2010/07/22 11:42 AM |
All FFT's need double precision - not | Vincent Diepeveen | 2010/07/24 11:39 PM |
All FFT's need double precision - not | slacker | 2010/07/25 03:27 AM |
All FFT's need double precision - not | Ricardo B | 2010/07/26 07:40 AM |
All FFT's need double precision - not | EduardoS | 2010/07/25 08:37 AM |
All FFT's need double precision - not | Michael S | 2010/07/25 10:43 AM |
All FFT's need double precision - not | Vincent Diepeveen | 2010/07/24 11:19 PM |
A bit off base | EduardoS | 2010/07/08 04:08 PM |
A bit off base | Groo | 2010/07/08 06:11 PM |
A bit off base | john mann | 2010/07/08 06:58 PM |
All right...let's cool it... | David Kanter | 2010/07/08 07:54 PM |
A bit off base | Vincent Diepeveen | 2010/07/19 03:36 PM |