AMD Adds MI450X And Publishes First MI455X Scores
The new Helios GPU is the flagship with four fewer memory stacks. The headline benchmark is AMD's own lab number, and AMD says so in the footnotes.
The new chip is the old chip with fewer stacks.
AMD added a smaller MI450X accelerator on November 12 and published its first measured MI455X inference score, 21,333 tokens per second.
The MI450X carries 288 gigabytes of HBM4 across eight stacks. The MI455X carries 432 gigabytes across twelve. Everything else matches: the same core compute, the same 23.3 terabytes per second of memory bandwidth, the same 40.3 petaflops of peak MXFP4 math and the same module shape for Helios racks. AMD didn't cut the chip. It cut the memory bill, by a third of the HBM4 in every rack.
The score belongs to the bigger part. AMD Performance Labs measured 21,333 tokens per second per GPU on DeepSeek-R1 in October 2026. AMD ran the test itself in an MLPerf Inference Offline configuration. It isn't an MLPerf submission, and the comparison target is Nvidia's published Vera Rubin result, entry 6.1-0106 in the Closed Division. On that basis AMD claims up to 30 percent higher throughput. A vendor's own lab number against a rival's filed number is a claim with a method attached, not a settled result.
The rest of the announcement filled in the family. The MI430X, aimed at scientific computing, carries 288 teraflops of native FP64 and the full 432 gigabytes of HBM4. A Helios rack in MI455X trim holds 72 GPUs, 31 terabytes of HBM4 and 1.67 petabytes per second of aggregate bandwidth. The MI450X product page lists 320 billion transistors on a mixed TSMC 2-nanometer and 3-nanometer process.
What's missing is the part buyers price. AMD lists no price for the MI450X and no first shipment date. The MI400 family launched on July 23, 2026 at the Advancing AI event, and the November update reads as a menu correction four months in: one rack GPU became two, split by memory capacity alone. Customers whose models fit inside 288 gigabytes get a cheaper rack. Customers whose models don't fit still pay for twelve stacks.
The audit trail is short. AMD's footnotes name the lab, the month, the configuration and the rival entry it measured against, and that's more disclosure than most launch claims carry. The number still needs an independent run before it prices anything. Until MLCommons holds an AMD-filed MI455X result, 21,333 tokens per second is a well-documented claim from the company selling the chip.
The conversation.
Loading comments…
A little more perspective.
Reading is free. Verified members can comment and reply for $2 a month.