GLM5.2 on AMD MI355X at 2626 tok/s/node at over 2x lower cost than Blackwell
Summary
The article reports on GLM-5.2 performance on AMD MI355X, achieving 2626 tokens per second per node at 2.4x rps, and claims over 2x lower cost than Blackwell. It details quantization to MXFP4 (vs FP8), the inference stack (sglang) and speculative decode optimizations, and MoE kernel tuning that yield substantial throughput gains. The piece emphasizes AMD software support as a key factor in attaining state-of-the-art throughput-per-dollar in frontier models.