Comment by quadrature

Not suprising that the hyperscalers will make this decision for inference and maybe even a large chunk of training. I wonder if it will spur nvidia to work on an inference only accelerator.