Request to bump species of AMD Radeon 780M

[arisu]

The 780M is currently species 4. I believe it should be increased to species 5. Despite being crippled by OpenCL, it consistently earns 400k PPD, and sometimes up to 800k PPD, even when it's sharing its TDP with a hungry CPU and its shader clock is limited to 75% maximum.

Its FP32 performance, 8.9 TFLOPS is almost twice that of the AMD Radeon Pro W5500, which is species 6 at 5.3 TFLOPS. Even if we consider its FP32 performance to be only half of the official specs, it's still faster than that of the slowest species 5 AMD GPU.

I know that the species is detected based on PCIe information and that it can't distinguish between a 780M, 760M, and 740M, but even the 740M is in the same league as another species 5 AMD GPU, the Radeon Pro Vega 20 (2.5 vs 3 TFLOPS). The 780M itself might even fit in species 6, but that would be too high for the 740M, so I'm not requesting that it be increased that high.

[muziqaz]

No

Yeah, nah, let me have a look into this, when I'm back home

[arisu]

I can do any benchmarks and tests on my 780M if you need. It's a pretty powerful GPU and runs circles (>4x more PPD) around GPUs that get much more work, even Nvidia GPUs. It never gets anywhere close to the timeout or deadline for any project thrown at it, even when running at 75% power.

[muziqaz]

No need for benchmarks.
You have 740m among those under same ID, that is the main argument against the bump.
But I'll have a look in the evening

[muziqaz]

Delete GPUs.json from your fah-client directory and restart fah-client. It should download new gpus.json
oh, what is your 780M's PCI ID?

[foxpy]

Delete GPUs.json from your fah-client directory and restart fah-client. It should download new gpus.json
oh, what is your 780M's PCI ID?

Wait, are you sure bumping 740M and 760M to species 7 was a good idea?

[arisu]

Delete GPUs.json from your fah-client directory and restart fah-client. It should download new gpus.json
oh, what is your 780M's PCI ID?

Vendor: 4098
Device: 5567

It's folding 16580 now and it seems to be doing well.

[muziqaz]

Delete GPUs.json from your fah-client directory and restart fah-client. It should download new gpus.json
oh, what is your 780M's PCI ID?

Wait, are you sure bumping 740M and 760M to species 7 was a good idea?

AMD species is not all performance based Species 7 might not be faster in some cases than high end Species 5, and that's ok

[foxpy]

AMD species is not all performance based Species 7 might not be faster in some cases than high end Species 5, and that's ok

To be honest, the entire concept of GPU species seems like a big mistery to me. Not sure if that's intended or not that they are so hard to research on for us curious donors

[muziqaz]

You won't be able to research, because it was done by one person
There is nothing to research. Species is just class of GPUs grouped by whatever one wants. nVidia species are more or less divided by generations. AMD species has a bit more thought put into that:
Species0: Unsupported
Species1: Terascale architecture (not supported anymore)
Species2: <1 TFLOPS
Species3: 1-1.999 TFLOPS
Species4: 2-4.999 TFLOPS (GCN Only)
Species5: 5+ TFLOPS (GCN Only)
Species6: RDNA1 Only
Species7: RDNA2 Low End (<6 TFLOPS)
Species8: RDNA2 High End
Species9: RDNA3
Species10: RDNA4
Species11: CDNA
Species12: Secret

It would have been much quicker to post a screenshot of those if my flickr worked

[foxpy]

You won't be able to research, because it was done by one person
There is nothing to research. Species is just class of GPUs grouped by whatever one wants. nVidia species are more or less divided by generations. AMD species has a bit more thought put into that:
Species0: Unsupported
Species1: Terascale architecture (not supported anymore)
Species2: <1 TFLOPS
Species3: 1-1.999 TFLOPS
Species4: 2-4.999 TFLOPS (GCN Only)
Species5: 5+ TFLOPS (GCN Only)
Species6: RDNA1 Only
Species7: RDNA2 Low End (<6 TFLOPS)
Species8: RDNA2 High End
Species9: RDNA3
Species10: RDNA4
Species11: CDNA
Species12: Secret

It would have been much quicker to post a screenshot of those if my flickr worked

Now that I think of it, shouldn't 780M be species 9 because it is RDNA3? Please don't see it as advice, I am just being curious.

[muziqaz]

Species 9 is for full blown desktop chips.
The only place 780M could fit was among low end RDNA2. If that particular PCI ID didn't have 740M included (thanks AMD, btw), 760M and 780M could have easily gone to species 9

[foxpy]

Yes, I agree, 740M sharing PCI IDs is very annoying

[arisu]

Species 9 is for full blown desktop chips.
The only place 780M could fit was among low end RDNA2. If that particular PCI ID didn't have 740M included (thanks AMD, btw), 760M and 780M could have easily gone to species 9

How come 5711 and 6400 weren't bumped to 9 then? Those are only 760M and 780M, without 740M. Unfortunately the only one I have is the 5567 one, which is 740M/760M/780M. At least now it's getting work.

[muziqaz]

They were