AMD shipped ROCm Core SDK 10.0.0 on 26 August 2026. The marketing date is the 27th. There is no ROCm 8. There is no ROCm 9. The previous production line was 7.14 in July.
That jump is not a CUDA-killer headline. It is AMD resetting the version number after a decade of HIP, putting the whole stack on TheRock, and shipping a new developer layer called ROCm.AI. If you serve models, the useful question is narrower: does this make Instinct boxes less of a science project than they were on ROCm 7.
What actually shipped
ROCm 1.0 landed in April 2016. Ten years later the same stack is supposed to cover Instinct accelerators, Radeon cards, and Ryzen integrated graphics on Linux and Windows from one pipeline. Windows retires the old HIP SDK and follows the ROCm Core SDK cadence. Minor releases are promised about every six weeks.
TheRock, which went production in 7.14, is now how 10.0 is built end to end: primitives, libraries, and framework wheels from one staged pipeline. AMD is also folding packages, the amdgpu driver, and public GPU tools toward a single repo.amd.com layout so you are not collecting keys from three hosts.
Validated inference frameworks in 10.0:
- PyTorch 2.13.0 (was 2.10.0)
- vLLM 0.27.0 (was 0.23.0)
- SGLang 0.5.15 (was 0.5.13)
- JAX 0.11.0 / 0.10.2
- TensorFlow 2.21
- MIGraphX 2.17
- ONNX Runtime 1.27.0
HIP picked up CUDA-parity pieces: hipMemGetDefaultMemPool, cooperative group inclusive/exclusive scans, faster hipEventRecord when timing is off. ROCprofiler-SDK can trace rocSHMEM and hipFile on the same timeline as kernels. That is operator tooling, not a model quality bump.
ROCm.AI is three products, not a compiler flag
The AMD newsroom and the ROCm 10.0 blog put ROCm.AI in front. It is not “ROCm 10 is 3.3x faster.” It is a workflow layer on top of the SDK.
AMD Skills. Validated ROCm knowledge in the Agent Skills format for Claude Code, Cursor, and Codex. Catalog: github.com/amd/skills. GPU skills walk an agent through rocm examine and standing up a vLLM OpenAI-compatible endpoint on MI300X / MI325X / MI350X / MI355X. CPU skills cover EPYC via ZenDNN / zentorch. This is documentation that an agent can execute, not a new kernel.
ROCm CLI. Tech preview. One binary for install, examine, serve, update. rocm serve <model> on PyTorch, rocm examine for driver/env. Works without a pre-existing ROCm install, including air-gapped bundles. AMD says it is version-agnostic starting at ROCm 7.13, with official 10 support coming later. Do not treat the CLI as GA.
Hyperloom. Open-source agentic loop: profile, plan, patch HIP/Triton/FlyDSL, benchmark, repeat. AMD says it runs on MI300X, MI325X, and MI355X and installs with pip. It is an optimizer you point at a serving stack, not a drop-in replacement for a known-good vLLM container.
The 3.3x number, without the press-release gloss
AMD Performance Labs, 7 July 2026, 8x Instinct MI355X:
- Inference: average 3.3x tokens/s vs ROCm 7.0, using a preview of ROCm.AI (ROCm 7.2.2 plus kernel, parallelism, and scheduling work) on GLM-5, Kimi-K2.5, and the DeepSeek R1 0528 checkpoint.
- Training: average 2.4x vs ROCm 7.0 with Megatron-LM on DeepSeek-V2-Lite, DeepSeek-V3-16B, and Qwen3-30B-A3B.
That is not “install 10.0, get 3.3x.” It is “a tuned ROCm.AI preview beat a ROCm 7.0 baseline on those models on that box.” Phoronix said the same thing. If your graph is Llama-70B on MI300X with last quarter’s vLLM wheel, you have no claim yet.
GIGAZINE also notes ComfyUI path work (Wan2.2, FLUX.2 Klein, SD 3.5 Medium). That is image/video serving, not chat completions.
What this means if you actually serve tokens
CUDA is still the default path because the kernel and serving ecosystem is denser. ROCm 10 is AMD arguing the gap is now packaging and agent workflow, not missing PyTorch.
For a cluster operator the checklist is boring and correct:
- Pin TheRock-built wheels. Do not mix 7.14 nightlies with 10.0 GA.
- Re-validate vLLM 0.27 / SGLang 0.5.15 on the exact Instinct SKU. MI355X numbers do not transfer to MI300X.
- Keep the ROCm CLI in a lab namespace until AMD marks ROCm 10 support, not “coming soon.”
- Treat Hyperloom as a perf engineer, not as CI that can rewrite kernels on prod.
- Windows teams: HIP SDK is gone. Core SDK or nothing.
Virtualization widened (passthrough and some SR-IOV on MI300X/MI325X/MI350P/MI355X). Partitioning did not change. OS matrix did not change. If you were blocked on Ubuntu/RHEL versions, 10.0 will not unblock you.
How IndieRouter thinks about it
IndieRouter serves OpenAI-compatible inference from GPUs in Indian data centres. The sovereignty question is where the weights and prompts sit, not which vendor logo is on the board. CUDA lock-in still decides most production graphs. ROCm 10 is the first AMD release in a while where “stand up vLLM on Instinct with a documented agent skill” is a real sentence instead of a forum thread.
We will not pretend AMD is a drop-in NVIDIA replacement on day one of 10.0. We will also not pretend the only legal inference stack is CUDA. If you are designing a DPDP-bound serving fleet, the interesting experiment is one Instinct node on ROCm 10 next to the CUDA fleet, same API, same eval set, honest tokens/watt. Not a press-release multiplier.
Docs: ROCm 10.0.0 release notes. Blog: ROCm 10.0 decade post.