
AMD Pitches Three Instinct Chips at Nvidia's Inference Lead

AMD has widened its Instinct MI400 line to three chips, pairing a scientific computing part with two inference options and claiming a lead over Nvidia on one of the benchmarks that matters most to buyers. The MI455X, running on AMD's Helios rack scale system, delivered 21,333 tokens per second per GPU on DeepSeek R1 in AMD's own testing.

And the comparison AMD draws is with Nvidia's Vera Rubin. The company says that throughput is up to 30 per cent higher than the latest published MLPerf offline result for Vera Rubin, and that Helios carries up to 20 per cent more memory bandwidth along with higher scale up bandwidth.
And the caveat is in the footnote. The figure comes from AMD Performance Labs, measured last month on an AMD tested MLPerf configuration, and set against Nvidia's published numbers rather than a run both companies submitted to the same audit.
But the portfolio split is the more practical news. The MI430X is aimed at simulation and AI for science, with native double precision compute, 432 gigabytes of HBM4 memory and up to 23.3 terabytes per second of bandwidth, and AMD published research across 47 workload configurations and 11 applications examining where emulating double precision is good enough and where scientific code needs the real thing.
The MI450X is the volume option, with 288 gigabytes of HBM4 per GPU and the same core compute, memory bandwidth and connectivity as the MI455X, so a customer whose model fits can buy the cheaper memory configuration and keep the same rack and software.
And that software point is AMD's quiet argument. Both chips run on the same Helios architecture and the same ROCm stack, which means a customer can move between them without re engineering a deployment, and Andrew Dieckmann, who runs AMD's data centre GPU business, framed the launch as giving customers options across precision, memory and configuration.
The competitive picture is the one to watch. Nvidia still sets the reference point that every accelerator is measured against, and AMD's claim rests on inference throughput on a single open model under a configuration it chose, which is the strongest ground it has held in this segment even if the margin is narrower than 30 per cent suggests.
MI455X claims 21,333 tokens per second per GPU on DeepSeek R1
AMD says that is up to 30 per cent above Nvidia Vera Rubin
MI430X carries 432 gigabytes of HBM4 for double precision work
MI450X offers 288 gigabytes and the same compute as MI455X
All comparisons are AMD tested against Nvidia's published figures
Source: AMD


