AMD has added a second GPU option to its Helios rack-scale AI platform and published new benchmark numbers that take direct aim at Nvidia's latest hardware. The company announced the Instinct MI450X on November 12, alongside measured inference results for the existing MI455X that AMD says beat Nvidia's Vera Rubin system by up to 30%.
The MI450X carries 288 GB of HBM4 memory per GPU and delivers the same core compute, memory bandwidth and scale-up and scale-out connectivity as the MI455X, AMD said. The idea is simple: customers building Helios systems can now match memory capacity to the size of the workloads they run, rather than paying for the MI455X's larger memory when they do not need it.
A benchmark aimed squarely at Vera Rubin
The headline number comes from AMD's own testing. In an MLPerf Inference Offline configuration, the MI455X running DeepSeek-R1 inside a Helios rack recorded 21,333 tokens per second per GPU, AMD said — up to 30% higher throughput than the published MLPerf Inference v6.1 result for Nvidia's Vera Rubin NVL72 system.
The comparison deserves a careful read. The figures come from AMD Performance Labs testing conducted in October 2026, not from an independent MLPerf-certified submission, and AMD compared its own result against Nvidia's published MLCommons entry, retrieved on November 11. Results will vary with model, configuration and software, and a single benchmark does not settle which platform is faster across all workloads.
Even so, the market liked what it heard. AMD shares rose more than 5% on Thursday after the announcement, trading around $585 in afternoon trading. Investors have been watching AMD's data-centre business closely: the division's sales surged in the third quarter, and the company has been winning Helios customers as cloud providers look for alternatives to Nvidia hardware.
What Helios customers get
Helios is AMD's rack-scale system — dozens of GPUs linked together in a single rack to run very large AI models — built on the company's ROCm software stack. The MI450X gives customers a second GPU option for the same rack, aimed at workloads where keeping running costs down matters more than maximum memory.
AMD did not announce a price for the MI450X, or say when systems built with it will ship. The announcement introduces the chip as an option for Helios customers, and the company says the new configuration was created in response to customer demand for an AI GPU with less HBM4 memory.
The wider AI chip race
The MI450X arrives as the AI chip race intensifies. Nvidia's Vera Rubin platform is its newest data-centre system, and AMD has been positioning the MI400 series — launched in July — as its answer. Anthropic said this week its own AI moved Claude onto AMD chips in three days, and Microsoft has chosen Helios for Azure AI workloads, giving AMD's platform real-world credibility beyond the benchmark slides.
For now, the MI450X is a choice, not a revolution: the same Helios rack, the same connections, less memory for customers who do not need the top tier. Whether it wins orders will depend less on AMD's benchmark claims and more on price, availability and how quickly customers can actually buy it — details AMD has yet to share.



