Tag
Engineers successfully serve GLM 5.2 on AMD MI355X at 2626 tok/s per node and 213 tok/s single stream, achieving ~80% of B200 throughput at over 2x lower cost than Blackwell.