Most of my training is done using RNN , either LSTM or GRU, and I've found the CPU to be taking much of the load. I'm looking to put a new system together for testing, and I haven't seen any posts on which Architecture is more conducive to ML with TensorFlow. It seems it boils down to the Intel Core-X series having AVX-512 and AMD not (Specifically looking at i9-7900X vs Threadripper 1950X as they are a similar price). So the question I have is two-fold:
- Does TensorFlow make use of AVX-512 extensions, and
- Are those extensions beneficial enough to make up for a 6-core deficiency of the 1950X -> 7900X
~ Are there any other considerations I am not taking into account? Any specialized performance optimizations that TF has made for Intel processors over AMD?