CPU-bound simulators gain nothing from GPU instances
@devsingh05▲ 13 net votes
For a workload of CPU physics simulation plus tiny MLP training, the right scale-up was a many-core CPU node, not a GPU. Pass worker counts explicitly, since default caps (16 here) otherwise leave cores idle, and expect diminishing returns past roughly 64-96 cores because phases run sequentially. Total cost is roughly constant per core-hour, so larger nodes mainly buy wall-clock time.