Summary
AMD reports 21,333 tokens per second per MI455X GPU on DeepSeek-R1 in an MLPerf Inference Offline configuration. It has also introduced the MI450X, a Helios GPU option with 288 GB of HBM4 memory.
AMD announced two updates to its Instinct MI400 Series on November 12, 2026: a company-reported inference result for the MI455X GPU running in its Helios rack-scale system, and a new MI450X GPU option for Helios deployments. AMD says the MI450X has 288 GB of HBM4 memory and shares the MI455X’s core compute, memory bandwidth and connectivity capabilities.
MI455X result measures offline throughput
AMD Performance Labs reported that an MI455X GPU in an AMD Helios system delivered 21,333 tokens per second on DeepSeek-R1 using an MLPerf Inference Offline configuration. AMD says this is up to 30% higher throughput than the latest published MLPerf offline result for Nvidia’s Vera Rubin NVL72 system in the Closed Division.
The comparison uses AMD’s October 2026 test result and the published MLPerf Inference v6.1 result for Nvidia, which AMD says it retrieved on November 11, 2026. Tokens per second measures output throughput in the benchmark; the Offline scenario is throughput-oriented and is not a measure of how quickly an individual user receives a response in an interactive session. AMD notes that results can vary with system configuration, model and software versions.
AMD also attributes the result in part to Helios system characteristics. It says Helios offers up to 20% higher HBM bandwidth than Nvidia Vera Rubin, as well as higher scale-up bandwidth. The company’s comparisons are based on published specifications and its own calculations.
MI450X adds a memory-capacity option
The MI450X gives Helios customers another GPU choice for matching accelerator memory capacity to a workload. AMD specifies 288 GB of HBM4 memory per MI450X and says it has the same core compute, memory bandwidth and scale-up and scale-out connectivity capabilities as the MI455X.
GPU memory holds model data and other information needed during inference. Capacity can therefore affect which models, context lengths or concurrent requests fit on an accelerator. AMD positions the MI450X for workloads that fit its memory capacity, while describing the MI455X as offering more headroom for larger models, longer contexts or higher concurrency.
AMD says customers can select either GPU within the same open Helios rack-scale architecture and ROCm software foundation. The added model expands the configuration choice for AI deployments without changing the platform’s stated compute, bandwidth or connectivity capabilities.