AMD says MI455X beats Nvidia's Vera Rubin by up to 30% on AI inference — but the test is its own

Advanced Micro Devices (AMD) said on Thursday (November 12) that its MI455X chip processed 21,333 tokens per second per GPU in DeepSeek-R1 inference testing, up to 30% above Nvidia's latest published MLPerf result for its Vera Rubin system. The company also introduced a lower-memory MI450X variant as it pushes deeper into the AI data-centre market.

AMD's own lab test, not an independent comparison

The figures come from testing conducted by AMD Performance Labs in October, while the Nvidia result was retrieved by AMD from the MLPerf website on November 11 (MLPerf Inference v6.1, Closed Division) — not a head-to-head test run by an independent body. AMD also notes in its release that results can vary with the model, hardware configuration and software. In other words, the "30% faster" claim is AMD's comparison on its own chosen terms, and the market still awaits independent verification.

MI450X arrives with 288GB of memory

AMD also added the MI450X to its Helios rack-scale platform, with 288GB of HBM4 memory per GPU and the same core compute, memory bandwidth and chip-to-chip interconnect as the MI455X, letting customers match memory capacity to their workloads. The MI455X carries 432GB of HBM4 at up to 23.3 TB/s of bandwidth. AMD says Helios offers up to 20% higher HBM bandwidth than Vera Rubin, along with higher scale-up bandwidth.

Shares rise about 6% as Nvidia changes chief the same day

AMD shares rose about 6% on Thursday after the announcement. Notably, Nvidia also had a major leadership change the same day: founder Jensen Huang stepped down as chief executive, president and director, with chief financial officer Colette Kress named interim chief executive (reported separately on this site). AMD launched the MI400 series in July, and Helios — its 72-GPU rack-scale design — is a direct answer to Nvidia's Vera Rubin NVL72. Sources: AMD newsroom, MLPerf.