Does GPU have to support double precision (FP64) to do folding?

[Scout]

hello, I just tried 2 laptops with these GPUs: Intel HD Graphics 620 and Intel UHD Graphics 730.
FAH client 8.3 could recognize them but shows "support: false". And I cannot download any WUs to fold on GPU.
Then I tried Nvidia T4 and Nvidia GTX4090, they worked just fine.
I searched in the forum and found that FAH uses double precision (FP64) data format. Nvidia T4 and GTX4090 supports it but few GPUs from Intel support it.
Just want to make sure, does GPU have to support double precision (FP64) to do folding?

Thanks a lot!

[calxalot]

Yes. Hardware FP64 is required.

[Joe_H]

FP64 started to be required about 6-7 years ago as some calculations in the GPU processing started to be done in double precision to carry enough accuracy. Not all calculations are done in double precision, the researchers identified a subset where the extra precision was required. Most calculations are done in single precision.

As for Intel graphics, if you go to the Wikipedia article listing their iGPUs and GPUs - https://en.wikipedia.org/wiki/List_of_I ... units#Gen9 - the iGPUs starting with what is identified as Gen 9 were the first ones recommended for and enabled for GPU folding. The lowest end chips were not fast enough to be useful though. There were enough small GPU projects 3-4 years ago that the rest did get assignments. I have no idea why there is no Gen 10, but Gen 11 iGPUs also were enabled for folding. Then with the Intel Xe / Arc "Gen 12" iGPUs and GPUs Intel made a decision to remove hardware FP64 support and move it into the driver. Tests of that software support by internal testers for F@h found that was not stable enough, and those have not been enabled for folding. The current Intel Xe 2 Battlemage chips have FP64 support returned to hardware, those have been enabled and are being tested now.

[Scout]

FP64 started to be required about 6-7 years ago as some calculations in the GPU processing started to be done in double precision to carry enough accuracy. Not all calculations are done in double precision, the researchers identified a subset where the extra precision was required. Most calculations are done in single precision.

As for Intel graphics, if you go to the Wikipedia article listing their iGPUs and GPUs - https://en.wikipedia.org/wiki/List_of_I ... units#Gen9 - the iGPUs starting with what is identified as Gen 9 were the first ones recommended for and enabled for GPU folding. The lowest end chips were not fast enough to be useful though. There were enough small GPU projects 3-4 years ago that the rest did get assignments. I have no idea why there is no Gen 10, but Gen 11 iGPUs also were enabled for folding. Then with the Intel Xe / Arc "Gen 12" iGPUs and GPUs Intel made a decision to remove hardware FP64 support and move it into the driver. Tests of that software support by internal testers for F@h found that was not stable enough, and those have not been enabled for folding. The current Intel Xe 2 Battlemage chips have FP64 support returned to hardware, those have been enabled and are being tested now.

Thanks very much!
Currently I have 2 laptops with Intel Iris Xe graphics and Intel HD Graphics 620, and one PC with Intel UHD graphics 730.
I tested UHD graphics 730 performance with Clpeak. Result shows that its FP64 performance is nearly the same with the PC's CPU intel i5-11400. Also, Nvidia T4 performance on FP64 is not much better than i5-11400. But I guess FP64 support at hardware level could help a lot on overall folding performance.
I will continue to test the performance of UHD 730, i5-11400 and Nvidia T4 with FAH Benchmark.

[Joe_H]

I am not surprised that the FP64 performance is similar on the UHD Graphics 730 to an i5-11nnn series CPU. As listed it is a Gen 12 chip, FP64 is provided through the driver instead of hardware on the iGPU. I haven't seen exact details, but that may be through submitting double precision calculations to the CPU connected to the iGPU. The other option would be emulating FP64 by using 2 FP32 arithmetic units together. But the real issue is that in testing the double precision provided through the driver was not stable and the WUs "blew up".

The HD Graphics 620 is a Gen 9 iGPU and is enabled in the GPUs.txt file. Potentially it could be used, but as a low end iGPU there appears to be little or no projects currently available small enough to utilize it for processing WUs.

[Scout]

I am not surprised that the FP64 performance is similar on the UHD Graphics 730 to an i5-11nnn series CPU. As listed it is a Gen 12 chip, FP64 is provided through the driver instead of hardware on the iGPU. I haven't seen exact details, but that may be through submitting double precision calculations to the CPU connected to the iGPU. The other option would be emulating FP64 by using 2 FP32 arithmetic units together. But the real issue is that in testing the double precision provided through the driver was not stable and the WUs "blew up".

The HD Graphics 620 is a Gen 9 iGPU and is enabled in the GPUs.txt file. Potentially it could be used, but as a low end iGPU there appears to be little or no projects currently available small enough to utilize it for processing WUs.

Thanks! One question: does FAH client use FP64 to do folding all the time, or it uses FP32 most of time, only uses FP64 when needed? Thanks!

[Joe_H]

I mentioned that in my first post, most calculations are single precision FP32 calculations with some double precision FP64 as needed. The actual calculations for folding are done in folding cores downloaded by the client and executing in their own processes. The client portion of the app installed handles downloads, uploads, and controls starting and stopping the folding core working on the WUs.

[Scout]

I mentioned that in my first post, most calculations are single precision FP32 calculations with some double precision FP64 as needed. The actual calculations for folding are done in folding cores downloaded by the client and executing in their own processes. The client portion of the app installed handles downloads, uploads, and controls starting and stopping the folding core working on the WUs.

Thanks a lot!