Comment by andrewSC

Am I missing something here or is inference going to be painful given the "low" memory bandwidth compared to, say, HBM2E?