> I’d be curious if this would just be a linear slowdown proportional to the RAM MHz or whether CAS latency plays into it as well.

Per o3-mini, the blocked gemm (matrix multiply) operations have very good locality and therefore MT/s should matter much more than CAS latency.