FP32 Performance on 90HX

[lholliger]

Is there any current method for getting FP32 performance to semi-normal? I'm trying to do Folding@Home with some cards I picked up (A nice space heater when I currently don't pay for electricity!) however it seems due to the noted limitation on FP32 performance they perform terribly compared to a regular RTX 3080 I have on hand. Currently running Fedora Linux however if FP32 can be fixed in a windows-only environment I can switch. At least from reading some other issues on this repo it looks like it could be a driver limitation and/or possibly a physical board mod?

nvidia-smi:

[bah86]

This is an unresolved issue.

[Demianvan]

Not yet resolved, but that limit is only over Float, then, if you have a way of "parse" your operations on Integers, the full power is used. In example, using IA models like Quantized LLM_INT8.

Look my benchmarks.

#73 (comment)