Or memory bandwidth (800GB/sec).

Or iGPU perf, which of course needs the above mentioned bandwidth.

Or ML inference performance (5 tokens/sec with llama 65B).