Upcoming OpenMM version doubles fahcore17 performance

[muziqaz]

In the latest Vijay Pande blog post here we can see the good news. 'Good' is a bit of the understatement
I hope the performance increase is across the board of various GPUs, not just high end ones. And when you think that if nVidia took their hands out of their a***es with opencl, this news would be even more exciting.
Hopefully new code won't take long to test out

P.S. I did not know where to post this 'news' so I dropped it here

Mod Edit: Moved Thread - PantherX

[Jesse_V]

Wow!

The post mentions a 2x increase in PPD, but won't the benchmark machine also get a speedup as well?

Simple economics drives Nvidia to support their CUDA more than OpenCL.

[k1wi]

It would sound like the benchmark machine won't. Perhaps because its optimisations on the GPU that don't translate for a reason to a CPU of the benchmark machines generation.

[P5-133XL]

Can the standard OpenMM cores also be recompiled with 5.1 and if so, would they get a similar speedup?

[bruce]

Wow!

The post mentions a 2x increase in PPD, but won't the benchmark machine also get a speedup as well?

Ordinarily this would be true but in this case, the benchmark machine runs SMP, not the GPU code. OpenMM isn't used on CPUs, only GPUs.

[Evil Penguin]

Can the standard OpenMM cores also be recompiled with 5.1 and if so, would they get a similar speedup?

You mean with core 16?
I doubt they would spend time doing so.

[artoar_11]

More details about OpenMM 5.0-5.0.1:

https://simtk.org/home/openmm - OpenMMProject Overview
https://simtk.org/news/news_details.php ... ews_flag=0 - OpenMM Project News
http://wiki.simtk.org/openmm/BenchmarkOpenMMDHRF - Benchmark: OpenMM on Dihydrofolate Reductase (DHFR)
https://simtk.org/forums/viewtopic.php?f=161&t=4054 - Current State of OpenMM on Kepler

[kiore]

From the blog:

The results have been pretty exciting. With OpenMM 5.1 (vs OpenMM 5.0, used in the current core 17 release), we are getting about a 2x speed up on typical FAH WU calculations, which will lead to an automatic 2x increase in PPD once this software is out of beta testing and integrated into core 17.
<snip>
There's a lot of testing to do and it's very possible that these numbers will change, but the results were so exciting that I wanted to give donors a heads up.

A doubling of speed is exciting, Prof Pande is excited.. and so am I

[PantherX]

If a jump from 5.0 to 5.1 is slightly more than twice the performance, I wonder what the performance gain would be from 5.1 to 6.0

[folding_hoomer]

The speedup is shown using the examples of a GTX680 and a HD7970.
I hope it will speedup also the "older" HD5- and HD6-series . . .
A HD 5870 performing 25K PPD - awesome

[bruce]

The speedup is shown using the examples of a GTX680 and a HD7970.
I hope it will speedup also the "older" HD5- and HD6-series . . .
A HD 5870 performing 25K PPD - awesome

I understand your concern ... and share the same concern, to an extent. Until we get actual measurements with the updated FahCores, it's all conjecture. Nevertheless, my conjecture is that you'll see a similar speedup in Core_17.

The charts in the blog show different amounts of improvement depending both on the hardware and on the type of calculation. Nothing can change that. Nevertheless, the changes are in OpenMM, not in OpenCL or in the drivers. That means the data and code being passed from OpenMM to OpenCL is more compact or more efficient. I can confidently predict that there will be improvements for HD5000 and HD6000 hardware, though it might not be the same 10%.

[k1wi]

What I found interesting was the comparison between the CUDA and OpenCL performance on the C2070. CUDA generally starts off stronger on a single GPU between (-2 and 12%) but OpenCL generally scales better.

[mdk777]

The speedup is shown using the examples of a GTX680 and a HD7970.
I hope it will speedup also the "older" HD5- and HD6-series . . .
A HD 5870 performing 25K PPD - awesome

I have to disagree with Bruce.

The new AMD architecture (GCN) was designed to maximized compute in general and open cl performance in specific.
The HD5870 was a previous generation, where AMD was still designing for graphics performance predominantly, and compute only ancillary.

So, while you may see some improvement, it will not be on a similar scale (2x) that you see with the new (GCN) cards.

[bruce]

Ok. We can disagree.

I think we're talking about two different things, though. OpenMM 5.0 with the old OpenCL drivers vs. OpenMM5.1 with the new OpenCL drivers is two different changes and each one will improve performance independently with perhaps a third factor by using the OpenCL API differently. I interpret what Dr. Pande is saying as changing ONLY the OpenMM version (which may or may not be what he said) and you're talking about changes to the drivers.

[mdk777]

I interpret what Dr. Pande is saying as changing ONLY the OpenMM version (which may or may not be what he said) and you're talking about changes to the drivers.

No, I just expect the hardware to respond differently to the same change in software. While the OpenMM version will see a (2x) speed in improvement on hardware fully optimized to run opencl; I expect some, but not necessarily the same response from hardware not optimized for opencl.

we might just be talking past each other, but my prediction is that a HD 5870 will not see the same percentage speedup as a AMD 7970.