---
format: "aidr-story-markdown/v1"
id: "741e662a29d0bd2102c968275be40545f3592758c606fd1f12e21c2045387c86"
canonical_url: "https://aidr.today/741e662a?lang=en"
title: "AMD MI355X Hits Record 1,129 Output Tokens Per Second on MiniMax M3.1 Flash"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-11-17T20:07:08.000Z"
category: "Infra"
topics: ["inference","framework"]
source_urls: ["https://marketbrief.now/ai/amd-mi355x-hits-record-1129-output-tokens-per-second-on-minimax-m31-flas-9ee50771","https://huggingnews.com/ai/amd-mi355x-hits-record-1129-output-tokens-per-second-on-minimax-m31-flas-9ee50771"]
summary: "vLLM and Inferact increased the performance of AMD MI355X GPUs on the MiniMax M3.1 Flash model by 4x in less than 20 days. The optimization achieved 1,129 output tokens per second (tps) at P90 during real production traffic. The gains came from new AMD kernels written in FlyDSL. These kernels are deterministic, meaning they do not use atomic adds to accumulate their outputs."
---

# AMD MI355X Hits Record 1,129 Output Tokens Per Second on MiniMax M3\.1 Flash

> [Open the canonical story](<https://aidr.today/741e662a?lang=en>)

**Published:** 2026-11-17T20:07:08.000Z
**Category:** Infra
**Topics:** inference, framework

## Summary

vLLM and Inferact increased the performance of AMD MI355X GPUs on the MiniMax M3\.1 Flash model by 4x in less than 20 days\. The optimization achieved 1,129 output tokens per second \(tps\) at P90 during real production traffic\. The gains came from new AMD kernels written in FlyDSL\. These kernels are deterministic, meaning they do not use atomic adds to accumulate their outputs\.

## Sources

- [Story source](<https://marketbrief.now/ai/amd-mi355x-hits-record-1129-output-tokens-per-second-on-minimax-m31-flas-9ee50771>)
- [Story source](<https://huggingnews.com/ai/amd-mi355x-hits-record-1129-output-tokens-per-second-on-minimax-m31-flas-9ee50771>)

