Cerebras' wafer-scale processor is claimed to move model weights 2,500 times faster than a GPU during LLM decoding

A post sharing Cerebras CEO Andrew Feldman's podcast discussion describes LLM inference as prompt processing followed by generating tokens one at a time. It says Cerebras keeps model weights in SRAM across its wafer-scale processor, while GPUs fetch them from HBM, and claims moving those weights into compute during decoding is about 2,500 times faster.