I've been using MuJoCo on Intel CPUs only. But recent releases from AMD look really cool if one wants to parallelize simulators. I'm curious if MuJoCo performs similarly on new AMD CPUs vs an Intel. Specifically, does the current MuJoCo release use any Intel-only optimizations? Any ideas would be greatly appreciated!
That is a very good question. In the documentation it states: gcc for linux and msvc for windows. So I assume that at least concerning compilation there are no icc-like intel-only optimizations. Another way I could guess would be if a matrix library such as MKL. But (fortunately !!!), I don't think so t.b.h. I think custom SIMD instructions are used which should be brand independent. Anyway, (unfortunately), I have never gotten my hands on a threadripper, so I cannot really make any all too useful statements.
Thanks for sharing your insights! Yeah if Intel intrinsics were used it's probably better to avoid AMD CPUs. It looks this is probably the case, the testings shown on MuJoCo website were done with Intel CPUs.