I agree with this, but I would offer that #3 is more an integration of #1 and #2 rather than a "learn", as your last sentence points out with "once you have don

I agree with this, but I would offer that #3 is more an integration of #1 and #2 rather than a "learn", as your last sentence points out with "once you have done 1) and 2)". So I believe you should more strongly recommend 1) and 2) and include what a later poster added about various frameworks: "BLAS, LAPACK, and maybe the whole CUDA family (CUDA, cuBLAS, cuDNN, etc)", see @madenine.