Hacker News Reader

Comment by semi-extrinsic | Hacker News Reader

This looks like a really stupid benchmark, no? 4 million cells in a standard aerodynamic simulation isn't enough work to even saturate the host CPUs on these systems, never mind the GPUs which have far "worse" roofline models from a CFD algorithm point of view.

I would not be surprised at all if these benchmarks ran faster if you removed the GPUs completely.

View on HN

That's the smallest of 4 experiments. It goes upto 140 million cells, where MI300X retains similar performance advantage of around 10% over Nvidia's H100.

Yeah, but the lack of comparison to a CPU-only system is suspicious. (As is the exaggerated plot without ticks and labels on the y-axis.)

I've never seen a non-reactive incompressible flow simulation get substantial speedup on GPUs. There are well understood fundamental reasons why this is the case.

I helped develop a hydro solver coupled with radiation at Los Alamos this summer. We observed from 7x upto 15x performance improvement on a single GPU compared to 64-core intel CPU. [1]

Now granted, the flops to byte ratio for this program might be better than an avg fluid simulator. Also, our performance tanked when we moved to multi-node system. But I am aware of underlying reasons behind the scalibility issues and they don't feel like problems that can't be overcome.

[1] https://github.com/lanl/HARD/tree/main

Nice, interesting solver! But yes, compressible/density-based solvers work a lot better on GPUs than incompressible/pressure-based solvers. When 80% of the runtime is spent solving a Poisson equation, GPU is not worth it.

Especially if you do the comparison on equivalent cost basis, i.e. "what is the walltime difference if I run on a $60k all-CPU cluster versus a $60k GPU cluster". Or in terms of cloud compute cost / HPC allocation spend.

This has not been my experience in the academic/research side. Poison solver-based incompressible CFD regularly runs ~10x faster on equivalently-priced GPU systems, and has been doing so since I've been following it (since 2008). Some FFT-based solvers don't weak scale ideally, but that'd be even worse for CPU-based versions, as they use similar algorithms and would be spread over many more nodes.

Reply on news.ycombinator.com