GLM 5.2 on AMD MI355X: half the price of Blackwell at 2,626 tok/s per node

Wafer has published the results of running the GLM 5.2 model on AMD Instinct MI355X accelerators. According to the reported figures, the configuration reached a performance of 2,626 tokens per second per compute node.

Of particular interest is the economic angle: according to the authors of the publication, the resulting processing cost at this configuration is more than two times lower than that of solutions built on Nvidia Blackwell GPUs.

Community response

The piece was published on Wafer's company blog and has already drawn the attention of the tech community — on Hacker News the news gathered 114 points and prompted 30 comments by the time this note was published.

The publication fits into a broader trend of recent months: inference of large language models is increasingly viewed as an area where AMD accelerators can compete head-to-head with Nvidia solutions not only on performance, but also on total cost of ownership.

The details of the testing methodology, the cluster configuration and the specifics of the GLM 5.2 model itself are not disclosed in the brief summary — readers interested in the technical details should turn to the original source.

Source: Hacker News