A new benchmark making the rounds among tech readers claims that AMD's latest AI accelerator can run a large language model faster and far cheaper than Nvidia's flagship hardware.
According to a post published on wafer.ai, the GLM5.2 language model running on AMD's MI355X chip reached 2,626 tokens per second per node. "Tokens" are the small chunks of text an AI model reads and generates, so tokens per second is a rough measure of how quickly a model can produce output. A "node" refers to a single server unit of the hardware.
The headline claim is about price as much as speed. The wafer.ai post states that this AMD setup delivered that performance at more than two times lower cost than Nvidia's Blackwell platform, currently the industry's dominant chip line for running and training AI systems.
The story surfaced on the front page of Hacker News, a widely read forum among software developers and technologists, where it drew 100 points and 25 comments — a signal that the claim is being actively discussed and scrutinized by a technical audience.
The figures come from a single source and have not been independently verified here, so readers should treat the specific numbers as claims rather than confirmed results.
Why it matters: Nvidia currently dominates the market for AI chips, and its hardware costs are a major expense for companies building AI products; credible evidence that a rival like AMD can match or beat that performance for less money could reshape how much the AI boom ultimately costs.