AMD’s New Chip Takes Direct Aim at Nvidia’s Vera Rubin

Human Verified 0

Share to: , and

Engineer in a cleanroom suit holding a silicon wafer in a semiconductor fabrication plant

AMD has introduced a new data centre GPU, the Instinct MI450X, and claimed its flagship MI455X beats Nvidia’s Vera Rubin on AI inference, in a direct challenge to the market leader. AMD said the MI455X delivered 21,333 tokens per second per GPU running the DeepSeek-R1 model, up to 30% higher throughput than Vera Rubin’s latest published result. The MI450X gives customers the same compute and connectivity with 288 gigabytes of memory per GPU.

What AMD announced

The announcement covers three parts of the MI400 series. The MI450X is a second GPU option for AMD’s Helios systems, with the same core compute, memory bandwidth and scale-up and scale-out connectivity as the MI455X, but with 288 gigabytes of HBM4 memory per GPU instead of the MI455X’s larger allocation. Customers can pick the chip that matches their memory needs.

AMD also shared new measured inference results for the MI455X on Helios, and new details on the MI430X, a chip aimed at scientific computing with native double-precision compute, 432 gigabytes of HBM4 memory and up to 23.3 terabytes per second of memory bandwidth. The MI430X targets researchers running large simulations alongside AI workloads.

The 30% claim, and what it is based on

The headline number comes with an important qualification: the test was run by AMD itself, not by an independent body. AMD’s labs ran the MI455X in an MLPerf Inference Offline configuration and compared it with Nvidia’s published MLPerf v6.1 result for the Vera Rubin NVL72, retrieved from MLCommons on 11 November.

AMD also claims Helios offers up to 20% higher memory bandwidth and higher scale-up bandwidth than Vera Rubin, based on its own calculations from published specifications. None of this is an independent finding that AMD’s chips are faster on all workloads — it is AMD’s case, argued with its own numbers.

Why now

The timing is deliberate. Nvidia’s chief executive resigned on 12 November, the same day AMD published its announcement, and the chipmaker’s interim leadership faces questions about its roadmap and its $500 billion financing platform. AMD is moving while its rival is distracted — and while customers are openly shopping for a second source of AI compute.

AMD’s own momentum helps. Its data centre business surged in the third quarter, and the MI450X adds a cheaper memory option to that story, for customers whose models fit within 288 gigabytes per GPU — a practical choice, not a flagship flex.

Server hall with rows of black rack cabinets and glowing blue status lights

Our Read

The 30% figure will travel further than the chip itself, and it should be read as a claim, not a verdict. AMD ran the test, picked the benchmark and chose the comparison — and it did so against a rival whose leadership is in turmoil. That does not make the number false.

What matters more is the MI450X itself. A second memory option for Helios, with the same compute and connectivity, is a practical answer to a real customer problem: not every workload needs the biggest memory configuration, and the cheaper option can win deals. Fractile’s $720 million raise this week shows investors are funding exactly this kind of alternative.

The wider picture is a chip race that is no longer a one-company story. Cerebras turned a surprise profit this week, Broadcom closed a $63 billion deal to supply Anthropic, and AMD is now claiming benchmark leadership. Nvidia still sets the pace — but the challengers are no longer asking permission.

Why It Matters

It matters because customers want a second source. Data centre operators have spent two years dependent on a single supplier for the most powerful AI chips, and every credible alternative weakens that dependence. If AMD’s numbers hold up in independent testing, the pricing of AI compute changes — and so does the balance of power in the industry.

It also matters for Nvidia. A benchmark challenge from AMD, in the week its chief executive resigned, adds pressure on the interim leadership to show that Vera Rubin and the roadmap behind it are still on track. The company’s response — in products, not statements — is now the thing to watch.

Background

AMD announced the MI400 series updates on Thursday 12 November, the same day Nvidia said Jensen Huang had resigned as president, chief executive and director, effective 11 November. AMD’s Instinct line has grown into the main challenger to Nvidia’s data centre GPUs, with Helios as its rack-scale answer to Nvidia’s NVL72 systems.

AMD gave no price or ship date for the MI450X, and no customer names. The announcement introduces the chip as an option for Helios customers; availability and pricing will follow when systems are ready, and the real test will be what customers report when they run it.

Source: AMD newsroom.

Share to: , and

Faheem R. Khattak

Senior automation engineer and AI marketing journo covering LLMs, algorithms, analytics, core updates, Bing search, and technical SEO.

0 Comments

Leave a Reply

Optional Take Tag:

🛡️ Spam protection: Honeypot + Time-Trap Active

This site uses Akismet to reduce spam. Learn how your comment data is processed.