The Pulse
AMD adds MI450X accelerator and reports MI455X inference results
AMD introduced the MI450X GPU for its Helios rack-scale system, offering 288 GB of HBM4 memory per accelerator. The company also reported new MI455X inference test results against a published Nvidia benchmark.
AMD has introduced an MI450X accelerator for its Helios rack-scale system, giving customers an alternative to the MI455X with a different memory capacity. In a company post on November 12, AMD also reported that an MI455X test produced 21,333 tokens per second per GPU on the DeepSeek-R1 model.
The company said the result represented up to 30% higher throughput than Nvidia’s latest published Vera Rubin MLPerf offline result. AMD’s figure comes from its own testing using an MLPerf Inference Offline configuration.
A second Helios option
The MI450X has 288 GB of HBM4 memory per GPU and the same core compute, memory bandwidth, and scale-up and scale-out connectivity capabilities as the MI455X. Both GPUs use the same Helios rack-scale architecture and ROCm software foundation.
AMD positions the new option for workloads that fit within its memory capacity, with throughput and response-time targets in mind. The MI455X provides additional memory headroom for larger models, longer contexts or higher concurrency, the company said.
How the inference comparison was made
AMD Performance Labs conducted the MI455X testing in October 2026. The Nvidia comparison used the published MLPerf Inference v6.1 Closed Division result for Vera Rubin NVL72, retrieved on November 11, 2026.
The company cautioned that system manufacturers can use different configurations and that results can vary with the model, configuration, software versions and other factors. Its reported throughput is a result for the stated DeepSeek-R1 offline configuration.
AMD also claimed that Helios has up to 20% higher HBM bandwidth and higher scale-up bandwidth than Nvidia Vera Rubin. Those comparisons were based on AMD Performance Labs calculations using published specifications; the memory-bandwidth calculations were made in November 2026, and the scale-up calculation was made in June 2026.
The rack-scale system
The MI400 series and Helios were launched at AMD’s Advancing AI event in July 2026. At that event, AMD described a Helios configuration with 72 MI455X GPUs and 18 sixth-generation EPYC “Venice” CPUs, connected through Pensando networking and supported by ROCm software.
AMD said in July that Helios was in production for planned deployment by AI companies at gigawatt scale. The new MI450X extends the GPU choices for Helios deployments rather than introducing a separate rack-scale architecture.