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

AMD launched ROCm 10 on August 27th and led with a number: an average 3.3× inference improvement and 2.4× training improvement over ROCm 7, on the same hardware. Then it printed the test conditions underneath, which is more than most vendors do, and the test conditions do not describe ROCm 10.

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 7.0 versus a 7.2.2 preview, run on 7 July, six weeks before the product being announced existed.

That alone would be pedantry. What follows is not.

So look at the serving stacks on either side of the comparison. Kimi-K2.5 was measured on vLLM 0.16.0 in the baseline and vLLM 0.22.0 in the new run. DeepSeek-R1 ran on an SGLang container built on 15 September 2025 against one built on 12 June 2026. Nine months of upstream inference work sits inside that gap, none of it AMD's, and the release attributes the whole delta to "AI-driven optimization of kernels, memory management and scheduling."

How much of 3.3× is the compiler and how much is vLLM? Nobody can say from these tables, and AMD is the only party who could have separated them by holding the engine version fixed. Doing so is cheap. Not doing so is a choice.

And 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-R1 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.

All of which is a shame, because the actual product is more interesting than the number.

ROCm Hyperloom is an agentic optimiser that profiles a workload, finds the bottleneck, writes the change, benchmarks it and checks correctness, targeting HIP, Triton and FlyDSL, with production support for vLLM and SGLang. AMD Skills ships curated AMD knowledge into Claude Code, Codex and Cursor through their own marketplaces, with each skill passing structural and behavioural tests before release. And the ROCm CLI — a technology preview, on Windows as well as Linux — installs as a prebuilt binary with no existing ROCm on the machine, manages side-by-side runtimes with rollback, and serves models from one 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, and if you have ever lost an afternoon to that, a version-agnostic CLI that installs and rolls back runtimes is worth more to you than any multiplier on a slide.

We would put the honest same-engine gain between a ROCm 7-era and a ROCm 10 stack on MI355X at well under 1.5×, and we would expect an independent benchmark to land somewhere near that once someone runs it with the serving version held fixed. If a third party reproduces 3.3× under those conditions, we will have been unfair and will say so.

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