Two numbers, two kinds of evidence
The comparison is better built than most vendor benchmark claims. Both sides run the same model, DeepSeek-R1, in the same benchmark family, MLPerf Inference, in its offline scenario. AMD names the exact Nvidia entry it compared against, 6.1-0106 in the v6.1 Closed Division, and dates its retrieval. A reader can find that entry and check the Nvidia side in full.
The AMD side cannot be checked the same way. Footnote MI400-031a of the release describes "an AMD-tested MLPerf Inference Offline configuration", run by AMD Performance Labs in October 2026. The release identifies AMD's figure as an AMD-tested configuration and Nvidia's as a published MLPerf result; it does not identify an MLPerf submission for the AMD run. MLPerf's value as a common yardstick comes from published submissions with disclosed systems and review. Until the AMD run appears there, or an independent lab repeats it on a shipping Helios system, the up to 30% figure rests on the vendor's own test.
None of this makes the number wrong. It sets its weight. Published submissions provide a documented comparison point, though buyers still need to examine configuration and workload applicability. A vendor-tested result needs its standing attached every time the number travels.
The release's other comparisons carry the same structure in a quieter form. AMD also cites up to 20% higher HBM bandwidth and higher scale-up bandwidth than Vera Rubin. Its own footnotes (MI400-004a and MI400-013) describe those as calculations from published specifications, made in November and June 2026 respectively, not as measurements of running hardware.
The GPU underneath the headline
AMD presents the MI450X as a lower-memory option with the same stated core compute, bandwidth and connectivity capabilities as the MI455X, on the same Helios rack architecture. The stated logic is fit: workloads that stay inside 288GB per GPU can take the lower-memory option, while larger models, longer contexts and higher concurrency stay on the MI455X, which the release describes as providing the additional memory headroom.
What the release does not give a Helios customer is the two facts a purchase plan needs. The release provides neither pricing nor an ordering or delivery timetable for the MI450X. As of November 12, the option exists in a portfolio announcement; its commercial terms are not in the release.
The scientific sibling
The same release fills in the MI430X, the MI400 variant for simulation and science, with FP64 double-precision support, 432GB of HBM4 and up to 23.3 TB/s of bandwidth. AMD pairs it with research examining where FP64 emulation is sufficient and where scientific codes need native double precision, across 47 workload configurations from 11 applications and benchmarks. AMD's editor's note says the MI430X information was added after the post was originally published. It is a narrower story than the inference claim, and for HPC buyers it may be the more durable one, because it rests on a capability difference rather than a vendor-reported comparison.
What to watch
Useful next evidence would include a published MI455X benchmark submission with configuration details, which would put its 21,333 tokens per second per GPU in the same record as the Nvidia result it cites, and MI450X pricing and delivery information. Until one of them appears, file the release as a vendor-reported measurement attached to a real product option, with the commercial terms not yet stated.
Sources
- [1] AMD, "AMD Advances the Instinct MI400 Series for AI and Scientific Computing" (primary release, November 12, 2026; benchmark methodology in footnote MI400-031a, bandwidth calculations in MI400-004a and MI400-013)Read source
- [2] AMD release as republished in full, including footnotes, Engineering.com, "AMD expands Instinct MI400 Series with new GPU options" (November 2026)Read source