# AMD MI355X runs GLM5.2 at 2,626 tokens/sec for under half Blackwell's cost

A benchmark shows AMD's MI355X accelerator achieving 2,626 tokens per second per node running the GLM5.2 language model, at more than double the cost efficiency of Nvidia's Blackwell platform. This demonstrates AMD's growing competitiveness in AI inference, potentially lowering costs for large-scale deployments.

**Importance:** 3/5

## Sources

### Technology
- [Hacker News](https://www.wafer.ai/blog/glm52-amd) — Fri, 03 Jul 2026 21:49:06 +0000