AMD has introduced the Instinct MI450X, adding another GPU option for its Helios rack-scale AI systems. The November 12 announcement pairs that expansion with new inference results for the MI455X.
The MI450X includes 288 GB of HBM4 memory per GPU. AMD says it retains the MI455X’s core compute, memory bandwidth and connectivity capabilities, while the MI455X offers more memory for larger models, longer contexts or heavier concurrency. Both use the same Helios architecture and ROCm software foundation.
AMD reports that MI455X delivered 21,333 tokens per second per GPU on DeepSeek-R1 in an AMD-tested MLPerf Inference Offline configuration, claiming up to 30% higher throughput than Nvidia’s published Vera Rubin result. The comparison comes from AMD’s labs; outcomes depend on hardware, software and configuration.
For AI infrastructure buyers, workload size determines which memory configuration fits their needs. The announcement does not specify pricing or a shipping date for MI450X.
