AMD Says Its MI400 Beats Nvidia’s Vera Rubin by 30% — the Claim, the Test, and the Timing

On 12 November 2026, the same day Jensen Huang resigned as chief executive of Nvidia, AMD chose its moment. The chipmaker announced new measured inference results for its Instinct MI455X graphics processor — and a new chip, the Instinct MI450X — in a direct challenge to Nvidia’s newest platform, Vera Rubin. The headline number is striking: AMD says its MI455X processed 21,333 tokens per second per GPU on the DeepSeek-R1 AI model, up to 30% higher throughput than Nvidia Vera Rubin’s latest published result on the same benchmark.

Quick answer: AMD claims its MI455X beats Nvidia's Vera Rubin by up to 30% on DeepSeek-R1 inference — but AMD ran the test itself. What the claim proves, and what it doesn't.

It is the most aggressive performance claim AMD has made against Nvidia in the AI era. It is also a claim that comes with a qualifier readers should understand before deciding what it proves: the test was run by AMD itself, not by an independent body. That does not make the claim false. It does mean the number deserves a closer look than the headline gives it.

Key takeaways

What AMD actually announced

The announcement, published in AMD’s newsroom on 12 November 2026, covers the company’s Instinct MI400 Series — its family of data-centre GPUs for AI and scientific computing — and it has three parts.

First, the benchmark. New measured inference results for the MI455X GPU on the AMD Helios rack-scale solution. Helios is AMD’s rack system: dozens of GPUs linked together in a single rack, designed to run very large AI models. Running DeepSeek-R1, a popular open AI model, in an MLPerf Inference Offline configuration, AMD recorded 21,333 tokens per second per GPU — up to 30% higher, AMD says, than the published MLPerf Inference v6.1 result for Nvidia’s Vera Rubin NVL72.

Second, a new chip. The Instinct MI450X gives customers building Helios systems a second GPU option alongside the MI455X, with 288 GB of HBM4 memory per GPU and — according to AMD — the same core compute, memory bandwidth and scale-up and scale-out connectivity as the MI455X. The MI455X carries 432 GB of HBM4, so the MI450X is the smaller-memory, presumably cheaper sibling for workloads that do not need the full memory headroom. AMD announced no price and no ship date for it.

Third, science. New details on the MI430X, a GPU aimed at high-performance computing and AI for science, with native double-precision (FP64) compute, 432 GB of HBM4 memory and up to 23.3 terabytes per second of memory bandwidth.

AMD also claims Helios itself has up to 20% higher HBM bandwidth and higher scale-up bandwidth than Vera Rubin, based on its own calculations from published specifications. That is a paper comparison, not a measured one, and AMD’s footnotes say so.

“AMD is advancing the AMD Instinct MI400 Series GPU portfolio with updates spanning scientific computing and AI,” the company said in its 12 November newsroom post, which also notes that the MI430X details were added after the post was first published.

The 30% number, unpacked

The 30% figure is the part of the announcement that will travel furthest, so it deserves the closest look. Here is exactly what sits behind it.

QuestionWhat AMD saysStatus
What was measured?DeepSeek-R1 inference, MLPerf Inference Offline configuration, MI455X GPU inside a Helios rackAMD’s own test, run by AMD Performance Labs in October 2026
What was the result?21,333 tokens per second per GPUA real measurement on AMD’s own system
What was it compared against?Nvidia Vera Rubin NVL72, MLPerf Inference v6.1, Closed Division, entry 6.1-0106, retrieved from MLCommons on 11 November 2026A real published result — not a head-to-head run on identical terms
How big is the gap?Up to 30% higher throughputAMD’s arithmetic on its own test versus Nvidia’s published result
Independently verified?No independent MLPerf-certified AMD submission exists yetUnverified by any third party

The comparison is legitimate in one sense: both numbers come from the same benchmark family — MLPerf Inference Offline, running the same model, DeepSeek-R1 — and MLPerf is the industry’s standard way of measuring AI inference speed. AMD’s footnotes cite the exact MLPerf entry and the date the figures were retrieved. That is more transparent than most vendor benchmark claims.

But it is still AMD’s own test, run by AMD’s own labs, on AMD’s own system, compared against a published Nvidia result rather than a head-to-head run on identical terms. Independent reporting of the announcement made exactly this point: the result is not an independent MLPerf-certified AMD submission, and it is not a finding that AMD’s chips are faster on all workloads. AMD’s own notes add that results may vary based on model, configuration, software and other factors.

What the test does — and does not — prove

What it does prove is narrower than the headline, and wider than nothing. AMD can run DeepSeek-R1 at 21,333 tokens per second per GPU on its own rack in its own configuration, and it can show its working against a named, published Nvidia result. That is a real measurement, and it is a credible signal that Helios is a serious system, not a paper launch.

What it does not prove is a 30% lead everywhere. One model, one benchmark configuration, one vendor’s lab. A 30% lead on DeepSeek-R1 inference does not automatically mean a 30% lead on other models, on training, or in real customer deployments — and AMD has not claimed otherwise. It also does not prove that Helios is shipping at scale: the MI450X has no price and no ship date, and the only named customer signal so far is cloud provider Vultr, which has said it will offer MI455X GPUs and support Helios, with preorders open for systems starting in the fourth quarter of 2026.

Why the timing matters

AMD did not pick a quiet day. The announcement landed on 12 November, the same day Jensen Huang stepped down as Nvidia’s chief executive after 33 years — the story we covered on 13 November. With Nvidia’s leadership in flux days before its fiscal third-quarter earnings, AMD pressed its opening.

The market noticed. AMD shares rose 5.4% on Thursday 12 November. But that reaction was to a benchmark claim, not an order book: no customer orders, no prices and no ship dates were announced alongside it.

The timing also sets up the week ahead. Nvidia reports its fiscal third-quarter results on Tuesday 17 November, with interim chief executive Colette Kress speaking on the call — the first of the two tests we flagged in our 14 November piece on the week ahead. If Nvidia’s numbers hold up, AMD’s claim becomes a footnote to a strong quarter. If they disappoint, the 30% number becomes the opening line of every story about the chip race.

What to watch next

Our take

Vendor benchmark claims deserve a three-question test, and it is worth applying here because the same test works on every chip announcement you will read this year.

Who ran it? AMD ran it, in AMD’s labs. That fails the independence test — for now.

Against what, exactly? A real published result, in the same benchmark family, on the same model, with the entry number cited. That passes the comparison test, and it is more than most vendor claims offer.

What is missing? A price, a ship date, an independent check, and any evidence beyond one model.

AMD has earned the headline. It has not yet earned the conclusion. The 30% number is a real measurement and a legitimate comparison — but it is a claim, not a verdict, and the verdict arrives when an independent body reproduces it or a customer buys the system at scale.

Frequently asked questions

What did AMD announce on 12 November 2026?

Three things: new measured inference results for the MI455X GPU on its Helios rack system (21,333 tokens per second per GPU on DeepSeek-R1, which AMD says is up to 30% higher throughput than Nvidia Vera Rubin NVL72’s latest published MLPerf result), a new MI450X GPU with 288 GB of HBM4 memory for Helios systems, and new details on the MI430X GPU for scientific computing.

Is AMD’s 30% claim independently verified?

No. The test was conducted by AMD Performance Labs in October 2026, and the comparison is against Nvidia’s published MLPerf Inference v6.1 result (Closed Division, entry 6.1-0106). It is not an independent MLPerf-certified AMD submission, and no third party has reproduced it yet.

What is the MI450X, and when can you buy it?

The MI450X is a new GPU option for AMD’s Helios rack-scale systems, with 288 GB of HBM4 memory per GPU and — according to AMD — the same core compute, memory bandwidth and connectivity as the larger MI455X. AMD has announced no price and no ship date for it.

What is the MI430X?

A GPU aimed at high-performance computing and AI for science, with native double-precision (FP64) compute, 432 GB of HBM4 memory and up to 23.3 terabytes per second of memory bandwidth. It rounds out the MI400 family for research workloads rather than commercial inference.

Why did AMD announce this on the day Jensen Huang resigned?

AMD has not said, and the announcement itself makes no mention of Nvidia’s leadership change. The timing is a fact — the announcement landed on 12 November, the same day Huang stepped down after 33 years — and the motive is interpretation. The market read it as a challenge: AMD shares rose 5.4% that day.

Sources and methodology

Method: every performance figure in this article is AMD’s own, labelled as such, and verified against AMD’s 12 November newsroom post and two independent reports published 13 November 2026. No independent reproduction of the benchmark existed at publication time. The article was desk-researched; the chips were not hands-on tested. Verification date: 14 November 2026.

Leave a Reply

Your email address will not be published. Required fields are marked *