CPU folding on iGPU?

[foldy]

I see. So the difference is the timeline or timeout. A slow iGPU cannot hold the deadline for GPU work units but it could help a CPU to complete CPU work units faster if a future FAHcore for CPU supports this.

[bruce]

Project deadlines are designed based on some guesses about how many processors are expected to run in parallel -- and, in effect, how many GFLOPS are expected. The total processing required can be adjusted within a range, too

The FAHCore is coded to use SSE2 hardware or AVX hardware (i86 or i64) or nV shader hardware or AMD shader hardware (OpenCL hardware). Code for an iGPU uses similar on-chip components that compete with code with FPU/SSE2/AVX operations. I think an iGPU would be a slow OpenCL device, not a fast CPU device.

[X-Wing]

I just ran a quick FAH Bench run on my system (Intel Core i3-8350K @ 4.0GHz, UHD Graphics 630, GTX 1660Ti) for a reference purpose, because this idea sounds really cool and I wanted to see if it was even feasible. My iGPU received a score of roughly 7.6 (I'm assuming this is in nanoseconds/day, for comparison purposes my 1660Ti typically makes 103-104 ns/day for around 700 kPPD), and my baseline CPU compute test got around 2.46 ns/day and my CPU folding yields around 30 kPPD. Based on this, it seems to me that the iGPU is around 2-3x faster than the CPU and has around 8% of my GPU's performance. I don't know if this helps or is in any way useful for this discussion, but the results also don't make much sense to me as I'm fairly new to FAH Bench, so if someone wouldn't mind explaining how to interpret these results more accurately, that'd be great.

[foldy]

It shows the iGPU is too slow compared to GPU folding. But if a new FAH CPU core would enable Intel iGPU acceleration then it would only speedup about 20% because it would only help the CPU. Developing a full GPU core for Intel iGPUs is too expensive for FAH currently because they focus on fast GPUs and CPUs.

[X-Wing]

Ok that makes sense, thanks for the clarification. I've read that Intel's Gen 11 iGPUs are supposed to come near and/or break the TFLOP barrier, which would put it nearly on par with my old GTX 750Ti that I used to use (it ran around 1.5 TFLOPS and 150 kPPD). Would that then be powerful enough to warrant development attention?

Here is the link to the article:
https://wccftech.com/intels-10nm-gen-11 ... g-in-2019/

[Joe_H]

When you ran FAHBench, did you get a report for its double precision performance as well? Or did it report that was unsupported? At least some support for double precision is necessary now for GPU folding. Whether or not the latest Intel iGPU's support double has not been reported so far that I have read.

[X-Wing]

I just ran the double precision benchmark, and received a score of 3.01 ns/day.

[bruce]

As far as double precision is concerned, the FAHBench speed is not particularly meaningful. When processing a GPU WU, FAH does require that the GPU is CAPABLE of running double precision (which yours is) but since it represents a tiny portion of a normal GPU WU, being faster or slower won't be particularly important when folding.

Since we don't have a clear picture of how an iGPU can potentially accelerate CPU folding, we have no idea IF or WHEN that might happen, but FAHBench doesn't shed any light on it.

[MeeLee]

I thought IGP could just run a GPU WU in parallel with the CPU?
Especially if it's supporting double precision.
I'm eager to see how those Intel NUCs with AMD graphics will perform.

[bruce]

Where do I see your system configuration. Have you posted it somewhere? Which AMD graphics device are you talking about?

[JimF]

The new Ryzen 3 3200G would probably be the first one I would fold on.
With 512 stream processors, clocked at 1,2Ghz, it wouldn't be strange to see scores of up to 100k PPD.

I typically saw 75k to 137k PPD on my i7-8700 (12 cores) running Ubuntu 18.04, depending on whether it was core a4 or a7. FAH seems to prefer Intel, but the Ryzen 3000 may change that. They have improved their floating point I believe.

[X-Wing]

What would it take to get that picture of how the iGPU could accelerate the CPU task?

[bruce]

What would it take to get that picture of how the iGPU could accelerate the CPU task?

Good question.

FAH doesn't deal with futures so you won't have information here until (and if) it is released and somebody reports actual results.

[X-Wing]

Sorry, I didn't word my question the best. I meant what, in terms of required programming knowledge/usage, would it take for someone to create/modify the necessary tools to get that picture? Like would the solution theoretically use GROMACS or OpenMM, in what language, etc.

[bruce]

FAH already uses GROMACS in FAHCore_a4 and FAHCore_a7, supporting CPU slots. Updating to a new version of GROMACS is a non-trivial development effort so they miss many new versions of GROMACS. New features that support new types of analysis are important; new features that improved performance are less critical. The update that supported AVX was important enough to create FAHCore_a7. The update to use relatively slow iGPUs (such as Intel's) are much less critical.

I can't make any predictions about how future Development funds will be spend but if (and when) this is incorporated will be a management decision.

The more powerful GPUs (generally excluding Intel branded devices) will likely be supportable running directly in a GPU slot.