AMD Adds a 288GB MI450X to the Instinct MI400 Line and Claims Up to 30 Percent Higher Inference Than Vera Rubin NVL72

AMD quietly widened its data-center GPU lineup on November 12, adding a second configuration to the Instinct MI400 family: the MI450X, which carries 288GB of HBM4 per GPU. It joins the flagship MI455X (432GB of HBM4) as the second GPU option for Helios, AMD's rack-scale system that integrates 72 GPUs per rack. The announcement also touched on the MI430X, a variant aimed at high-performance computing and scientific workloads.

The pitch for the MI450X is configurational rather than architectural. AMD says it carries the same core compute, memory bandwidth, and scale-up and scale-out connectivity as the MI455X — the only difference is memory capacity, so customers can match a configuration to their workload instead of paying for headroom they will not use. AMD gave no price and no shipping date for the new part.

Alongside the new GPU, AMD published inference results for the MI455X on Helios: 21,333 tokens per second per GPU running DeepSeek-R1 in an MLPerf Inference Offline configuration, tested by AMD Performance Labs in October. AMD claims that is up to 30 percent higher throughput than Nvidia's published MLPerf Inference v6.1 result for Vera Rubin NVL72. Two qualifiers matter here: the comparison is against Nvidia's published number retrieved from MLCommons on November 11 — not a head-to-head run under identical conditions — and every figure comes from AMD's own labs, so the claim is company-reported and has not been independently verified.

The timing is not accidental. The MI400 series launched in July 2026 as AMD's answer to Nvidia's rack-scale strategy, and the first gigawatt of the 6-gigawatt OpenAI agreement is built around MI450-series deployment. With Nvidia consumed by a leadership crisis this week, AMD is putting measured inference numbers and a mid-tier memory option on the table precisely as Vera Rubin begins to ramp.

There is a real product logic behind the 288GB variant, too. Not every inference deployment needs 432GB of HBM4 per GPU; serving smaller models or latency-sensitive workloads can favor cheaper memory configurations that preserve the same interconnect fabric. A two-tier lineup lets AMD price into accounts that would otherwise default to Nvidia's lower-memory parts.

What happens next is the part AMD cannot control. Vendor-published MLPerf results have a long history of flattering the vendor that runs them, and the true test of the MI450X and MI455X will be independent measurements and, more importantly, production deployments at scale — starting with whether the OpenAI build-outs actually deliver on schedule. Until then, the 30 percent claim is a target painted on Nvidia's door, not a settled fact.