Pretty impressive what the Cerebras folks seem to have accomplished here, reportedly verified by Argonne National Labs. Interested in hardware folks’ take on the architecture as a realistic path forward for training larger and larger models.