AMD Adds MI450X With 288GB HBM4, Posts MI455X Inference Results

What happened

The MI450X is a second GPU option for the Helios rack-scale system. AMD says it carries the same core compute, memory bandwidth and scale-up and scale-out connectivity as the MI455X, differing in memory capacity so customers can match a configuration to their workload. AMD gave no price or shipping date for the MI450X.

The new MI455X figures come from AMD Performance Labs testing in October: on the Helios system, running DeepSeek-R1 in an MLPerf Inference Offline configuration, the GPU delivered 21,333 tokens per second per GPU. AMD says that is up to 30% higher throughput than Nvidia’s published MLPerf Inference v6.1 result for Vera Rubin NVL72.

How the benchmark was run

The 30% comparison carries a qualifier: the test was run by AMD’s own labs, and the comparison is against Nvidia’s published result on MLCommons, not a head-to-head run under identical conditions. AMD notes the Nvidia figure was retrieved from MLCommons on Nov 11, and that results may vary with hardware, software and configuration.

The MI400 Series is AMD’s data-center GPU family launched in July 2026; the flagship MI455X carries 432GB of HBM4 per GPU, and the Helios rack integrates 72 GPUs. The announcement also covers the MI430X for high-performance computing and scientific workloads.

Key facts

1. MI450X: New GPU option with 288GB of HBM4 per GPU, the same core compute, bandwidth and connectivity as the MI455X; no price or shipping date given
2. MI455X: DeepSeek-R1 inference of 21,333 tokens per second per GPU on Helios, in an MLPerf Inference Offline configuration tested by AMD Performance Labs in October
3. Comparison: AMD says up to 30% higher than Nvidia Vera Rubin NVL72’s published MLPerf v6.1 result, a comparison against published data, not a head-to-head run
4. Helios: AMD’s rack-scale system integrating 72 MI455X GPUs per rack
5. MI430X: Another model in the announcement, aimed at high-performance computing and scientific workloads