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