GPU Benchmarks? WU/GPU mismatching

[cine.chris]

Based on Mr Chodera`s GitHub comment

"Just a heads-up that we're working on extensive benchmarks of a variety of systems to be able to gather data to be able to predict performance on individual GPUs based on some easily-computed features of any project."

Three questions and a comment:

Wondering if this info will be available? Or perhaps incorporated into FAHbench?

Also wondering if any of the Geekbench openCL results are meaningful for Folding?

Looking at recent Proj 13416 results, the GTX cards were thriving, the RTX cards were struggling.
1660s 1100137
1660ti 1108804
2060s 848627 1135460
2070s 1252834
2080ti 1754242
Frustrating...
When the 2080ti was at 4.7M PPD on a different project a few hours prior, but nice that the 1660s are at 140% of their usual 7-800K.
It would be nice to see a more ideal matchup between WU/GPU.
I think the frustration I feel leads to user abandonment of Folding, even now, understanding that atom#s effect GPU utilization.

[bruce]

Looking at recent Proj 13416 results, the GTX cards were thriving, the RTX cards were struggling.
1660s 1100137 TU116 (1408)
1660ti 1108804 TU116 (1536)
2060s 848627 1135460 TU106/TU104 (1920-2176)
2070s 1252834 TU106/TU104 (2304-2560)
2080ti 1754242 TU102 (4352)
Frustrating...
I think the frustration I feel leads to user abandonment of Folding, even now, understanding that atom#s effect GPU utilization.

That much is common knowledge, but it's a lot deeper than that. e.g.- What about AMD GPUs?

Yes, it looks like having more that about 1400 shaders doesn't help much with the projects that are currently active.

[foldy]

One unofficial option is to run 2 work units concurrently on the same GPU to use all shaders.

[cine.chris]

Guess I asked too many questions, as the first one regarding benchmark results status didn`t get any attention.

One unofficial option is to run 2 work units concurrently on the same GPU to use all shaders.

I`d be willing to test the double WU option. Is there a link to the how-to set this up?

[Ichbin3]

Is there a link to the how-to set this up?

It's inofficial, so what are you expecting ....
Use your own imagination.
One hint: it seems to be easier to do this at Linux.

[Ichbin3]

2080ti 1754242
Frustrating...

I can't complain about the 13416. Still getting between 2.85 and 3.2 Mill PPD with this kind @2080Ti@180W

[bruce]

Assuming that you successfully get two WUs processing concurrently, you're faced with the fact that they will NOT finish at the same time, so reverting to one WU is also something you'll need to figure out (unofficially).

[cine.chris]

I can't complain about the 13416. Still getting between 2.85 and 3.2 Mill PPD with this kind @2080Ti@180W

13416 was at 4.3M on the 2080Ti today, as I happened to be reading this.
Also saw my highest estimated peak output of 13.9M, all slots were loading,so, it`s not always a struggle.

[JohnChodera]

> Wondering if this info will be available? Or perhaps incorporated into FAHbench?

We've got a new data scientist helping us out for a few weeks, and we're hoping he'll have the bandwidth to dive into this in the next week or two and start refining GPUSpecies and compiling expectations of PPD for various cards to post publicly.

We're also hoping he can resurrect FAHBench, update it to match core22, automate the update process, and make sure everything is consistent between FAHBench and what you can expect the card to deliver in PPD.

~ John Chodera // MSKCC

[MeeLee]

Doubling/tripling doesn't always go well with smaller WUs, and could lead to increased GUI response time (lag).
From Boinc, most RTX GPUs seem to be able to pair 2 or 3 WUs just fine.
3 to 4 seems to induce lag, and past 4 WUs isn't really recommended, as the gui would show regular freezes of well over 1 second to sometimes almost 5 seconds (Linux).

I guess it'll depend on how much each WU calls on double precision or half precision, since the GPUs seem to perform best at single precision.

There already is quite an amount of data available on the forum on what GPUs are the fastest.

A good thing with slower WUs on core 22, means you can actually drag the power limit lower, and still gain good GPU boost frequencies, as they will process more like core 21 WUs.
If Core 21 and 22 WUs are getting closer to each other in terms of processing power;
one can use a single, more aggressive economy setting, on the power slider, and just add another GPU.

Then the question is, which GPU is better?
If core 22 WUs no longer benefit from 4000+ cores of 2080Tis, then perhaps purchasing a mid-range RTX (like a 2060 KO or Super, or a 2070 Super), will be more beneficial.
Either way, most of the 2070/2070 Supers work really well at almost the same TDP as the RTX 2060/2060 Super.

[cine.chris]

Either way, most of the 2070/2070 Supers work really well at almost the same TDP as the RTX 2060/2060 Super.

I likely won`t pursue multi-WU, preferring to KISS & focus on other projects.
Being a retired engr, I`ve been seeking solutions that balance PPD/cost & PPD/watt. For example, I`d likely choose to do a dual 2070super over a single 2080Ti for a system mostly dedicated to F@H.

Right now, I`m anxiously waiting to see what AMD & Nvidia will be able to deliver in the Fall.
AMD caught Intel napping (comatose actually) and I`ve built two AMD systems this year, so I`m a hardware agnostic. Currently finishing up my 12kBTU(1ton) mini-split heat-pump install in the basement, thus the operational environment won`t dictate what can be running.

It would be nice to have an efficient mix of GPU & WU working in a framework that helps the real-world, while allowing our simple-minded gamer egos to accumulate points and have a game strategy.

[bruce]

I guess it'll depend on how much each WU calls on double precision or half precision, since the GPUs seem to perform best at single precision.

There already is quite an amount of data available on the forum on what GPUs are the fastest.

FAH does not use ANY half-precision operations (FP16) That's used by ray-tracing and AI.

The use of double-precision (FP64) is significantly less that single precision (FP32). I have no information about how those percentages might vary depending on the nature of the specific analysis.