[gmx-users] AVX2 SIMD intrinsics speed boost

Erik Lindahl erik.lindahl at scilifelab.se
Fri Jul 12 09:48:37 CEST 2013


Hi,

We will have AVX2 acceleration ready for general usage before the end of July (together with some other goodies), and it will be markedly faster, but until it's ready and tested to give correct results we can't say anything about the final performance.

However, in general AVX2 will have the largest effects on the same parts of the code that run on the GPU when one is present, which means it might not provide a huge speedup when used in combination with accelerators.

Cheers,

E.

On Jul 11, 2013, at 6:01 PM, Bin Liu <fdusuperstring at gmail.com> wrote:

> Hi all,
> 
> If my understanding is correct, GROMACS parallelization and acceleration
> page indicates AVX2 SIMD intrinsics can offer a speed boost on a Haswell
> CPU. I was wondering how much performance gain we can expect from it. In
> another word, what's the approximate speed increase if we run a simulation
> with AVX2 SIMD intrinsics on a Haswell CPU (say i7 4770K) than on an Ivy
> Bridge CPU of the same  clock (say i7 3770K) with the current AVX SIMD
> intrinsics? And is there a timeline for the release of AVX2 SIMD intrinsics?
> 
> This information is crucial if we want to assemble a machine with balanced
> CPU and GPU performance.  My current machine has i7 3770K (3.5GHz, stock
> frequency) and Geforce 650 Ti (768 CUDA cores, 1032MHz). When I ran
> simulations with   rcoulomb=1.0 and rvdw=1.0, I got this at the end of the
> log file:
> 
> *Force evaluation time GPU/CPU: 1.762 ms/1.150 ms = 1.531*
> *
> *
> It seems I need a GPU with 50% more CUDA cores. In the best scenario, If
> AVX2 can give 30% speed boost, and I can successfully overclock 4770K to
> 4.5GHz, I need 1900 CUDA cores( 130%*(4.5GHz/3.5GHz)*1.531*768 cores) at
> the same frequency to get balanced CPU and GPU performance. Then I will
> need a GeForce GTX 780 (2304 CUDA cores at 863MHz, equivalent to 1925 CUDA
> cores at 1032MHz). Since GROMACS is highly insensitive to memory clock and
> latency, I hope this naive arithmetic can give a good estimation which
> graphic card I should purchase.
> 
> Best
> 
> Bin
> -- 
> gmx-users mailing list    gmx-users at gromacs.org
> http://lists.gromacs.org/mailman/listinfo/gmx-users
> * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/Search before posting!
> * Please don't post (un)subscribe requests to the list. Use the 
> www interface or send it to gmx-users-request at gromacs.org.
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists




More information about the gromacs.org_gmx-users mailing list