MiniMax M3.1 Flash Runs 4x Faster on AMD MI355X in vLLM Within 20 Days
AMD accelerators have often lost to Nvidia because of unfinished software, not weaker hardware. SemiAnalysis reports that MiniMax M3.1 Flash running in the vLLM engine on AMD MI355X got 4x faster in less than 20 days. The gain came from new Inferact kernels written in FlyDSL, which sum in a fixed order and do not use atomic addition. Anyone calculating token cost on AMD should recheck old measurements.
AI-processed from X: @SemiAnalysis_; edited by Hamidun News
When a team picks hardware for running language models, it often goes with Nvidia. The reason is not that the cards are always faster, but that the software for them has been refined for years. For AMD, the picture was the reverse: the accelerator is powerful, but the software lagged behind. That is why a short post from SemiAnalysis analysts drew attention: according to their data, the speed of MiniMax M3.1 Flash in the vLLM engine on the AMD MI355X accelerator grew 4x in less than 20 days.
What exactly happened
First, a quick decoder for the names. AMD MI355X is an AI accelerator, a counterpart to Nvidia's top cards. vLLM is an open-source engine that runs language models and serves requests to them. MiniMax M3.1 Flash is the model used for the measurement. Inferact is the company that released the new kernels for AMD cards. FlyDSL is the language those kernels are written in.
In GPU programming, a kernel is a small program for a basic operation, such as matrix multiplication or combining expert outputs in a MoE model. The quality of these programs directly determines how many tokens per second a card can produce. A poor kernel means an expensive card sits idle. According to SemiAnalysis, the 4x gain comes from the new Inferact kernels. It is a typical software optimization story: the card stayed the same, but it now runs noticeably faster.
An important caveat: the post gives no absolute numbers. There are no tokens per second, no starting baseline, and no test settings. So "4x" is growth relative to the starting point, not a verdict that "AMD is faster than Nvidia."
What "deterministic" kernels mean
The authors specifically note that the new kernels are deterministic: they never use atomic addition to accumulate results. Imagine a hundred cashiers all writing their takings into one shared cell at the same time. Atomic addition guarantees that no entry gets lost. But the order of the entries is slightly different every time.
In school arithmetic, the order of the terms doesn't matter. In floating-point computation, it does: because of rounding, the sum can differ in the last digits. That is why the same request to a model on the same hardware sometimes returns slightly different numbers from run to run. A deterministic kernel assembles the result in a fixed order, so the answer repeats bit for bit.
Here is where that comes in handy. Say you updated your vLLM version or changed settings and want to know whether the model has started answering worse. If the computations don't repeat, you can't tell real degradation from noise. With reproducible kernels, the test gives the same result every time. The same applies to debugging a strange answer and to reinforcement learning, where mismatches between generation and training corrupt the signal.
Reproducibility usually costs speed. That is why the combination is what makes this interesting: the kernels are deterministic and still delivered a 4x gain. The post does not claim, however, that one caused the other. It only reports that the gain was achieved with such kernels.
What to do in practice
First, don't write off AMD based on measurements from a month ago. If your token-cost calculation or cluster choice relies on old numbers, they may be out of date by a large factor. Second, test the result on your own model: the gain was shown for one specific model, MiniMax M3.1 Flash, and on other architectures it may be different. Third, if you need reproducibility, check which kernels are enabled in your vLLM build.
There are also fair limitations. This is a post on X, not a technical report: there is no methodology, no comparison with Nvidia, and no data on latency or price. The 20-day gain could have started from a low point, when support for the model had only just appeared. But the direction is clear: competition in inference is increasingly decided at the level of kernels and compilers, not just silicon. The better the open-source software for AMD gets, the more options teams have, and the more leverage in price negotiations.
- Nov 18, 2026Sciforium pushes DeepSeek-V4.1-Flash on AMD to 1 129 tokens per second
- Nov 18, 2026Hugging Face routes LLM requests through OpenRouter, cutting capacity errors by 46%
- Nov 18, 2026Nvidia's data-center revenue jumps 110% to $107.6 billion: the AI pause didn't hurt
- Nov 18, 2026BRIA releases RMBG-3.0: background removal in 6 seconds, beats BiRefNet in 75.9% of head-to-head comparisons
Key Questions Answered
What exactly happened?
First, a quick decoder for the names. AMD MI355X is an AI accelerator, a counterpart to Nvidia's top cards. vLLM is an open-source engine that runs language models and serves requests to them. MiniMax M3.1 Flash is the model used for the measurement. Inferact is the company that released the new kernels for AMD cards. FlyDSL is the language those kernels are written in. In GPU programming, a kernel is a small program for a basic operation, such as matrix multiplication or combining expert outputs in a MoE model.
What "deterministic" kernels mean?
The authors specifically note that the new kernels are deterministic: they never use atomic addition to accumulate results. Imagine a hundred cashiers all writing their takings into one shared cell at the same time. Atomic addition guarantees that no entry gets lost. But the order of the entries is slightly different every time. In school arithmetic, the order of the terms doesn't matter. In floating-point computation, it does: because of rounding, the sum can differ in the last digits.
What to do in practice?
First, don't write off AMD based on measurements from a month ago. If your token-cost calculation or cluster choice relies on old numbers, they may be out of date by a large factor. Second, test the result on your own model: the gain was shown for one specific model, MiniMax M3.1 Flash, and on other architectures it may be different. Third, if you need reproducibility, check which kernels are enabled in your vLLM build. There are also fair limitations.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.