Hybrid CPU architecture performance impact on RLlib and Tune

How severe does this issue affect your experience of using Ray?

  • None: Just asking a question out of curiosity

I’m looking to purchase a new CPU and am wondering about the impact of the hybrid architecture of the new Alder Lake CPU’s on using RLlib in combination with Ray Tune. These new Intel CPU’s contains both Big cores and Eco cores. (example: Intel i9-12900K).

The instruction set extensions supported by the Big cores and Eco cores are not the same. For example, the big cores support AVX512 but the Eco cores do not. (Intel has disabled AVX512 support in recent microcode updates, but let’s ignore this for the sake of the question here).

My question is, when running multiple RLlib trials with Ray Tune, using all available CPU cores to step through environments, would the presence of the Eco cores negatively impact the performance of the Big cores?

Specifically for the provided example: would the presence of the Eco cores (which don’t support AVX512) prevent the Big cores from using AVX512 when running RLlib environments in parallel?

Or in other words: would the supported CPU instruction-set extensions be limited to the ‘least common denominator’?

I’ll make my decision to go with Intel or AMD based on the answers to the above. So many thanks in advance!

Self-answering my question:

It seems like Intel has already done the work to reduce the instruction set to the least common denominator.

This link has all the answers: Intel's new Chimera: Alder Lake - Agner's CPU blog

1 Like