AMD says ROCm 10 is 3.3x faster than ROCm 7. Its own footnote says neither side ran ROCm 10

AMD printed the test conditions under its headline number, which is more than most vendors bother to do.

Then you read them.

ROCm 10 launched on 27 August with an average 3.3x inference improvement and 2.4x training improvement over ROCm 7, on the same hardware. Footnote MI350-81 says the measurement is an eight-GPU MI355X system running ROCm 7.0 against a similarly configured system running "a preview version of AMD ROCm.ai (ROCm 7.2.2 with optimizations such as Optimized Kernels, Parallelism and Scheduling)."

So it is one point release against a preview of another, measured on 7 July, six weeks before the product being announced existed.

That alone would be pedantry.

The serving stacks are nine months apart

Look at what actually ran on either side. Kimi was measured on vLLM v0.16.0 in the baseline and v0.22.0 in the new run. DeepSeek ran on an SGLang container built in September 2025 against one built in June 2026.

Nine months of upstream inference work sits inside that gap, and none of it is AMD's. The release attributes the whole delta to "AI-driven optimization of kernels, memory management and scheduling."

So how much of the speed-up is the compiler and how much is vLLM? Nobody can tell from these tables. AMD is the only party that could have separated the two, by holding the engine version fixed, and doing that is cheap. Not doing it is a choice.

The footnotes lose track of themselves

The training test is headed ROCm 7.2.2 and then lists "ROCm 7.2.1 + Primus v26.3" for all three models.

In the inference table, the DeepSeek baseline names a container image tagged sgl-dev-v0.5.2 while the version text beside it reads V0.5.13, which is the version of the run it is being compared against.

Somebody assembled this from two spreadsheets (we have all done it, just not in a launch release).

Which is a shame, because the product is better than the number

ROCm Hyperloom is an agentic optimiser. It profiles a workload, finds the bottleneck, implements the change, benchmarks the result and validates correctness, targeting HIP, Triton and FlyDSL, with support for vLLM and SGLang.

AMD Skills pushes curated AMD knowledge into Claude Code, Codex and Cursor through their own marketplaces, and each skill passes structural and behavioural testing before release.

And the ROCm CLI, a technology preview on Windows as well as Linux, installs as a prebuilt binary onto a machine with no ROCm on it at all, manages side-by-side runtimes with rollback, and serves models from a single interface.

That last one is the real fix

AMD's software reputation was never mainly a compiler problem. It was a which-container-do-I-run problem.

If you have ever lost an afternoon to that, a version-agnostic CLI that installs and rolls back runtimes is probably worth more to you than any multiplier on a slide.

We would put the honest same-engine gain between an old stack and a new one on that hardware at well under one and a half times, and we would expect an independent benchmark to land near there once somebody runs it with the serving version held fixed. If a third party reproduces the full claim under those conditions, we were unfair and will say so here.

Ten years of ROCm, and the thing most likely to win AMD developers this year is a command that installs it.