In one paragraph
ROCm is AMD’s open-source GPU computing platform: compilers, runtimes, and libraries for AI and other compute work, aimed at Instinct, Radeon, and Ryzen AI hardware, on Linux and Windows.[1]You do not install “ROCm” the way you install a chat app. You install amatched set: a kernel driver the GPU will talk to, a userspace that knows yourgfx ID, and (if you use them) framework wheels that were built for that userspace.
This site’s daily driver does not wait on that full stack. On our test rig the Qwen3.6-35B-A3B Unsloth UD-Q5_K_XL run is llama.cpp Vulkan onone Radeon AI PRO R9700 (gfx1201, 32 GB). ROCm is still the reason the rest of the AMD AI ecosystem exists.
Why it exists (CUDA vs ROCm)
Most machine-learning software was written against CUDA, NVIDIA’s GPU compute stack. If you buy a Radeon, that stack is not on the card. AMD’s answer is ROCm: an open platform with its own runtime, compiler, and libraries, plus HIP — a C++ API and kernel language that looks familiar to CUDA programmers and can be used to port CUDA code, without being a drop-in replacement.[2]
The name is literal. AMD’s November 2016 Supercomputing announcement called it theRadeon Open Compute Platform.[3]The pitch then was the same shape as now: an open GPU compute stack so you are not locked to one vendor’s closed toolchain. The distance from that press release to a working home LLM is the rest of this guide.
Go deeperHIP is not “CUDA but orange”
AMD’s own HIP page is explicit: GPU programmers with CUDA experience will find HIP straightforward, and HIPify can rewrite many CUDA calls, but HIP isnot intended as a drop-in CUDA replacement. Expect manual work and performance tuning.[2]
For a home user that means: a PyTorch ROCm wheel is not “the CUDA wheel with a different pip extra.” It is a different build, with a different gfx target list, and it will refuse to load kernels your card’s ISA is not in. That is why this site’sGPU directory leads with gfx IDs, not marketing names.
What it means for home users
Three practical facts, in the order they usually hit you:
- Your card has an ISA name. The Radeon AI PRO R9700 and the RX 9070 XT are both
gfx1201on AMD’s ROCm 10.1.0 hardware table (RDNA 4).[4]Wheels and llama.cpp HIP builds must list that target. The product page is the SKU; the gfx ID is what the compiler cares about.[6] - Official support is a table, not a vibe. “Official” on this site means the card or OS appears on AMD’s current compatibility matrix or the matching release notes. “Works” means a community or first-party run that is not on that table. We keep those words apart on the versions page and in thedirectory.
- You can generate tokens without finishing ROCm. llama.cpp’s Vulkan backend talks to the desktop GPU driver (RADV on Linux, the vendor ICD on Windows). That is how this site’s 35B daily driver runs. HIP/ROCm is what you need for PyTorch, vLLM, and the AMD-documented AI ecosystem.[5]The backend chooser starts at llama.cpp either way.
What this guide covers
Eight pages, written for a person who bought one Radeon and wants local Qwen:
- This overview.
- The stack — amdgpu and KFD, ROCr, HIP, LLVM/clang and comgr, libraries, frameworks.
- Architectures and gfx IDs — RDNA vs CDNA, why the target matters,
HSA_OVERRIDE_GFX_VERSION. - Vulkan vs HIP — when llama.cpp Vulkan wins, when HIP/ROCm wins.
- Versions and support — cadence, official vs works, distros, 10.1 and the R9700, the ROCm CLI.
- History — a cited timeline.
- Glossary.
- Misconceptions and FAQ.
Install steps live in the install tutorial, not here. When something breaks, use the error paste andtroubleshooting pages. Cited consumer gaps live onROCm’s open problems — discuss them on theforum with a rig attached. Fit a model with thecalculator before you buy a second card.
Sources
Every version and support claim on this page points here. If a sentence has no mark, treat it as lab opinion or a TODO — not as AMD policy.