AMD says MI355X passed 1 million tokens per second in MLPerf Inference 6.0

AMD said its Instinct MI355X GPUs passed 1 million tokens per second in the MLPerf Inference v6.0 round, released April 1, 2026, using multinode submissions. Its own table shows an 11-node Llama2-70B offline run at about 1.04 million tokens per second, against about 103,000 on a single node. These are self-submitted benchmark results, so they support the inference case for Instinct without proving production demand.

• Llama2-70B offline on 11 nodes: 1,042,110 tokens/s (MI355X)
• Llama2-70B offline on 1 node: 103,480 tokens/s
• GPT-OSS-120B offline on 12 nodes: 1,031,070 tokens/s
• MLPerf Inference v6.0 results released April 1, 2026