Page 1 of 1

CMP 90HX for non-FP32 workloads

Posted: Mon Dec 02, 2024 2:59 pm
by lukash
Hello,

I recently came across some NVIDIA CMP 90HX mining cards which are allegedly just 3080 dies with many functions disabled by fuse/firmware/driver, thought they would be good for folding however they were only getting 200-400K PPD, while my daily driver 3080 gets ~7M/day. Looking more into these cards I found they have massively reduced performance in non-integer loads, although seemingly less of a hit if there is less/no FP32 FMA calls or FP64, and then of course nearly no hit for integer loads since it's needed for mining.

From what I've seen somehow disabling FMA (Fused Multiply-Add) leads to mostly unlocked performance, however you then run into precision errors so it looks like that would be an issue. I would also think the core would need to be recompiled without FMA in order for this to work, which seems unlikely.

Are there any causes I could go for or other things I could change to work around these cards limitations?

Re: CMP 90HX for non-FP32 workloads

Posted: Mon Dec 02, 2024 9:00 pm
by toTOW
There's absolutely no chance that FAH will build a specific core just for them.

You'd better sell those cards and get a standard GPU.

Re: CMP 90HX for non-FP32 workloads

Posted: Mon Dec 02, 2024 10:08 pm
by lukash
toTOW wrote: Mon Dec 02, 2024 9:00 pm There's absolutely no chance that FAH will build a specific core just for them.

You'd better sell those cards and get a standard GPU.
That's understandable. Do all the simulations rely on FP32 FMA, or do some causes have less chances of running into this sort of math? From what I've heard almost all simulation is FP32 FMA-dependent, just hoping there's something else in case. I am trying to sell them though, the price was good enough where I could sell higher anyways.

Re: CMP 90HX for non-FP32 workloads

Posted: Wed Dec 04, 2024 8:51 pm
by toTOW
I don't know, but since FMA is a common function, I don't see any reason why it wouldn't be used.

Note that although most FAH computation are FP32, it also requires FP64 support for a few parts of the code.