Benchmark on a10g and h100

[AlexanderMath]

Amazing work. Thanks for the benchmarks on A100 and V100. Wondering if anyone tried a10g and h100s?

[aris1978]

Is double precision card faster than single precision card? Will the A100 be faster than the L40 for gpu4pyscf? If I want to purchase a new GPU, should I consider A100 or V100 over L40 or even the RTX4090/5090?

[wxj6000]

Currently, GPU4PySCF is running faster on A100 over L40 in general. For certain algorithms such as density fitting, A100 can be 10x faster than L40. But the consumer grade GPU (RTX4090/RTX5090) is much cheaper.

[aris1978]

Can GPU4PySCF runs on more than 1 GPU?

[wxj6000]

@aris1978 Yes, you can run it on a multi-GPU system. This feature is still in experimental. We are still in the progress of improving the performance.

[aris1978]

In the multi-gpu system, will the VRAM be additive? (for example, two GPUs with 16GB of VRAM will be able to run jobs that require 32GB of VRAM on a GPU)

[wxj6000]

@aris1978 Yes, the large intermediate variables will be distributed across multi-GPUs. This feature is still experimental. Please let us know if you find any issue.