Folding Forum

Which fah project number would that be?

Grandpa_01 wrote: More like tripple, the projects I run that use AVX generrally run about 3 time as fast as when AVX is not used on the same project.

I didn't know that there were any folding projects/cores using AVX. I think that newest GROMACS version is now AVX capable but I didn't know that any of the folding cores had been converted so as to use it. The NaCl core is by far the newest GROMACS core and I don't know if it even uses the AVX capabilities. Perhaps you are using the term "projects" in the more general (non-folding) definition.

P5-133XL wrote:

Grandpa_01 wrote: More like tripple, the projects I run that use AVX generrally run about 3 time as fast as when AVX is not used on the same project.

I didn't know that there were any folding projects/cores using AVX. I think that newest GROMACS version is now AVX capable but I didn't know that any of the folding cores had been converted so as to use it. The NaCl core is by far the newest GROMACS core and I don't know if it even uses the AVX capabilities. Perhaps you are using the term "projects" in the more general (non-folding) definition.

I did not say fah projects I said other projects, just stating what the actual difference is with other cpu crunching intensive work. The speed up with AVX is huge and is also limited by memory the faster the memory the greater the speed up. AVX floods the memory and the memory can create a bottle neck. If fah develops a AVX capable client/core there is quite a potential for a very large and real speed up. Other projects have already done this and the results are quite impressive.

Both SSE and AVX can provide excellent speed up of single precision floating point math. Most of the benefits of SSE could only be obtained with the help of an outstanding programmer. The current version of Gromacs is based on the work of one such programmer. AVX has more potential and a good compiler might do as well or even better, depending on many factors, but only relatively new processors and relatively new OSs have AVX support. Of course many are going to be comparing the speedup of AVX to ordinary x86 code and they can give more impressive numbers than if they were comparing AVX to code that is already highly optimized for SSE.

As long as FAH has to support Donors who do not have AVX support, the Fah infrastructure support will have to get more complex to support both. I don't have any realistic way to estimate what percentage of FAH donors still depend purely on SSE and how many have both AVX capable hardware running on an AVX capable operating system.

NaCl uses a version of GROMACS that supports AVX, however I don't know if it is enabled (I would guess not as it is probably a generic compiled exe). There are also a lot of other improvements and speedups in newer GROMACS that all systems can take advantage of. How much of a boost AVX would bring depends on how cache and memory bound folding is (a linpack would get most of the full theoretical boost). My guess would be a decent jump, which may vary between projects (similar to how the GPU projects seem to be variably performing).

To what I read most of the relevant OS support AXV2 and as for hardware Haswell upwards (AXV with SB upwards)

And the current BA was also made for a minority; my guess is that the base with AXV(2) capable systems is even bigger. Let's hope a "new" core 0xa7 or 0xa8 (wild guess on names) gives us that power.

Finding a good programmer to harvest that power is a challenge, I can understand.

The 2x speed up already mentioned was conservative and specific only to FAH, because FAH is already so highly optimized for SSE.

Any old timers will recall the doubling of fah speed when we moved from the Pentium 4 to the Core 2 Duo. C2Ds could process twice the SSE instructions per clock cycle. AVX does approximately the same thing for fah, eventually.

Old timers will also remember the ~3.5x speed-up when SSE optimizations were optional in Core_78 (out of a 4x theoretical maximum). That's when the real SSE programming work was added to FAH.

WRT speedups from AVX2: Native Code Performance on Modern CPUs: A Changing Landscape (slides).