We're actually talking about the difference between Cheng using 8 GPUs and 2 CPUs while Google used 16 GPUs and 40 CPUs. These are under-your-desk levels of resources. Cheng et al authors are all affiliated with UCSD which owns the Expanse supercomputer which is orders of magnitude larger than what you would need to reproduce the original work. Cheng et al does not explain why they used fewer resources.

That’s a fair complaint then.

No it's not. They ran it longer instead.

The 2022 paper pretty explicitly says that runtime is not a substitute. They say their best result "can only be achieved in our 8-GPU setup".

I assume you mean Fig. 6 here?[0]

But that was explicitly limited to 8 hours for all setups. Do they have another paper that shows that you can't increase the number of hours of a smaller GPU setup to compensate?

[0]https://dl.acm.org/doi/pdf/10.1145/3505170.3511478

They also changed the ratio of RL experience collectors to GPU workers (~1/20th the RL experience collectors, 1/2 the GPUs). I don't know what impact that has --- maybe each GPU episode has less experience? Maybe that makes for an effectively small batch size and therefore more chaotic training? But either way, why change things when you can just match them exactly?