> fully programmable Tensor Processor Cores Some of those statements are inaccurate. Being an infra/HPC dev, I have yet to see anyone programming custom kernel

> fully programmable Tensor Processor Cores Some of those statements are inaccurate. Being an infra/HPC dev, I have yet to see anyone programming custom kernels for Intel Gaudi. Even if they exist, Tensor Cores on either platform are far from "fully programmable". Even on CUDA: - It's only for tiled matrix multiplication... - Of dense matrices (structured sparsity doesn't count)... - Using only addition and multiplication, - For very few numeric types. Meaning that there are almost no applications outside of reimplementing BLAS.