Criteria for GPU retirement

[Roadpower]

So this question is just out of curiosity and boredom.

What makes for the reasons that the FAH team will disqualify or retire GPU's in the line up for acceptable processors? It is a time calculation? Just too old for current techniques or requirements? Very low percentage of the user base? In my guess thinking about this it makes sense to retire very old processors as they would likely start having a negative impact on efficiencies for the entire project, again just a guess.

[Neil-B]

If a GPU can't do the calculations the Cores require such as Double Precision Floating Point (FP64) and are incapable of running OpenCL to the at least a certain version 1.2 for the most part (possibly 1.1 for some edge cases) then it will be taken off the white list … If A GPU can do these but not meet the Expiration Deadline then it simply is not up to folding unfortunately (although I can't recall anyone posting their GPUs can meet this) so the limiting factor at the moment seems to be where the GPUs are actually able to run the folding core.

The use of Timeout (with QRB for completing within this) and Expiration date is the teams way of setting the expected standards - any GPU that can complete before Expiration is still deemed viable.

This is actually quite a lot harder for CPUs as the limiting factor is actually their speed and a low core count on an old "slow" CPU can really struggle to finish before the Expiration - made even worse if not 24/7 folding … I am aware of a good number of folders who have retired first their GPU because it couldn't meet the technical requirements and then much later their CPUs cause they were just too slow

Though there was a folder recently whose CPUs failed the technical capability test as they were 32bit Xeons from the early 2000s - so some do fail the technical capabilities hurdle rather than fail on being too slow.

[Roadpower]

Thanks @Neil-B

Appreciate your time to answer this.

[bruce]

At some point FAH decided that SSE enhancements to CPU folding was a good thing but it was optional. then FAH discovered that almost everybody had newer processor so someone made a formal decision to require it

The same sort of thing happened when somebody decided that NVidia GPUs needed to be Fermi or better. I suppose there was some feature that Fermi introduced but it was a really long time ago. The Double Precision requirement was more subtle. Small projects could run on Single Precision but bigger projects needed Double Precision to maintain adequate accuracy. Then somebody rewrote the code using Mixed Precision which maintained the accuracy with very little performance degradation (DP is much slower than SP, but that also depends on your hardware). Eventually there were no more projects being generated for Single Precision so those GPUs that couldn't support DP essentially just ran out of work. There was no formal announcement.

[JimboPalmer]

Anecdotes:

My HD 2600 Pro had a core written in a proprietary language called Brook. When AMD gave up supporting Brook for OpenCL, it could not be used anymore.
My G 205 and GT 440 used CUDA, when F@H switched to OpenCL they got much slower, then unsupported as F@H needed OpenCL 1.2.
If they still worked at all, they would be too slow to fold.

Currently I am running a GTX 1050ti, (140k PPD) a GTX 1060,(400k PPD) and a relatively new GTX 1650. (300k PPD)

The underlying science sets the needs.

For CPUs GROMACS still supports SSE2, (CPUs back to 2001) but uses avx_256 if the CPU supports it (2011) There are a avx2, and a avx_512 instruction sets, but they are not widely supported yet. (no AMD runs axv_512 yet)

https://en.wikipedia.org/wiki/GROMACS
https://en.wikipedia.org/wiki/Advanced_ ... Extensions

More importantly both graphics card drivers and OSes are starting to demand 64 bit CPUs. (2003)

On GPUs, three limits bite us:
OpenMM needs OpenCL 1.2 and it need Double Precision floating point math (FP64)
https://en.wikipedia.org/wiki/Molecular ... n_GPUs#API
https://en.wikipedia.org/wiki/OpenCL
The third limit is the speed of the card, most usually judged by the number of shaders/cores The GT 1030 has 384 cores and can complete Work Units if run 24/7. If your card has less than 256 cores I am skeptical it can complete a Work Unit in time, even if run 24/7.

Looking above my HD 2600 Pro had 120 cores, my G 205 had 16 cores, and my GT 440 had 96 cores so none would be finishing on time if if they could start work.

[Wild speculation begins here]

SSE2 is just slower than avx_256, I do not expect it to be dropped soon. It is more likely that CPUs will just become too slow.

Officially Fermi and Tesla GPUs by Nvidia do not support OpenCL 1.2. Tesla is unsupported now, but Fermi may become unsupported soon. Nvidia is not working on OpenCL 2.0 for Fermi.
https://en.wikipedia.org/wiki/Fermi_(microarchitecture)

On the AMD GPU side Terascale 2, Terascale 3 and GCN 1.0 only support OpenCL 1.2, if OpenCL 2.0 becomes a requirement, then these cards would not fold
https://en.wikipedia.org/wiki/TeraScale ... chitecture)
https://en.wikipedia.org/wiki/Graphics_Core_Next

[bruce]

I am aware of a good number of folders who have retired first their GPU because it couldn't meet the technical requirements and then much later their CPUs cause they were just too slow

I've got a system that folds on CPU only becaust the GT710 (192 cores) in it is just too slow to meet the deadlines. I should probably pull the GPU out of it and use the iGPU and save a couple of Watts. The CPU gently keeps the room from getting too cold in winter.

[Neil-B]

Wow .. all of a sudden slow GPUs come out of the woodwork ... apologies all I was wrong ... there is such a thing as a too slow GPU with 1.2 and FP64

[Roadpower]

Sorry for the late follow up. This turned into a very educational conversation. Thank you all