AMD says MI455X beats Nvidia on AI inference benchmark

AMD said new testing of its Instinct MI455X graphics processing unit, running on the company’s Helios rackscale system, produced 21,333 tokens per second per GPU on the DeepSeek-R1 model under an MLPerf Inference Offline configuration. AMD says that result tops Nvidia’s latest published Vera Rubin MLPerf offline score by up to 30%. The company also says Helios delivers up to 20% higher high-bandwidth-memory (HBM) bandwidth, and higher scale-up bandwidth, than Vera Rubin.

Alongside those results, AMD is expanding the Instinct MI400 Series with a new chip, the MI450X. It carries 288 GB of HBM4 memory per GPU, while matching the MI455X on core compute, memory bandwidth, and inter-GPU scale-up and scale-out networking. That gives Helios customers two memory options on the same open rack-scale architecture and AMD’s ROCm software stack: the MI455X for larger models, longer context windows or higher concurrency, and the MI450X for workloads whose memory needs fit a smaller footprint while still hitting throughput and response-time targets.

“AMD Instinct MI400 Series GPUs bring leadership performance and new capabilities to AI and scientific computing,” said Andrew Dieckmann, AMD’s corporate vice president and general manager for its Data Center GPU Business. “With new milestones in AI inference performance and an expanded range of configurations, we’re giving customers more options to power their next generation of AI workloads.”

AMD frames the expanded lineup as a response to the differing requirements of frontier AI training versus interactive inference, which the company says calls for different trade-offs among raw performance, memory capacity and response time.

Why it matters

AMD’s benchmark figures are self-reported and have not been independently verified, but they put a specific number behind AMD’s pitch against Nvidia’s next-generation Vera Rubin platform as both chipmakers compete for AI data-center GPU orders.