AMD, Intel disagreements 'red alert'

Hello, I have like 4 Intel Matlab clients and 1 powerful AMD Matlab client (Ryzen 1950, overclocked). I need to do heavy calculations (e.g. LSTM) with the AMD client, but still be able to reproduce the results on the other clients namely Intel (be aware that rand is resetted, this one is not the hurdle)

My problem is, it does not produce the same. All Intel agrees with each other, but not with the AMD and of course vice versa.

Is there any way to make AMD agree with Intel, or Intel agree with AMD processors with regards to LSTM or likely other Neural Network computations?

[dpb]

Down to the LSB level, probably not.

What kind of discrepancy are we talking about here?

[Martin]

With regards to the discrepancy: I get results between -1 to 1. On AMD I can easily get e.g. -0.2 and 0.3 on Intel...

Answer 1

[Ned Flanders]

You should test the workaround for AMD CPUs to use the AVX2 codepath. Its possible that with Intel using AVX2 and AMD using SSE codepaths, the results might somewhat differ for specific opperations. Also, AVX2 is much faster on your TR box.

See: https://www.reddit.com/r/matlab/comments/dxn38s/howto_force_matlab_to_use_a_fast_codepath_on_amd/