Ubuntu 24.04+ on bare metal is still the measured AMDai path —
/setup,
/gpus/r9700,
/benchmarks.
Everything on this page is Cited, not measured on the local desktop
unless a Windows amdai-bench/run@1 card lands.
None has. We do not invent drivers, version pins, tok/s, or prices. llama.cpp first.
AMD’s HIP SDK for Windows is a documented subset of ROCm (runtime, HIP APIs, tooling), not the full Linux userspace we measure on the local desktop. The 2026-07-31 Windows requirements table lists Radeon AI PRO R9700 and several consumer RDNA3/RDNA4 cards as Runtime + HIP SDK. For a first token, follow llama.cpp’s Windows Vulkan build docs: install the LunarG Vulkan SDK, then build with GGML_VULKAN=ON, or use a llama.cpp Windows x64 Vulkan release. HIP/ROCm on Windows is second, and only after Vulkan tokens work on a card AMD lists.
Do not start here
WSL2 — technical preview, not a first-run path
ROCm 10.1.0 release notes add WSL2 support in TheRock artifacts as a technical preview: Ubuntu .deb packages and Python wheels, not RPM. Profiling, debugging, and KFD-dependent tools are out of scope at that launch. AMD SMI on WSL2 is discovery/monitoring only. The notes also list hangs and missing device-side fault reporting. That is too thin to recommend WSL for a first Windows token. The cited WSL walk is /setup/wsl. Prefer Ubuntu 24.04+ on bare metal if you want the measured AMDai cards to match the box.
Zero to first token
1 · Cited, not measured on the local desktop
See the discrete GPU
Confirm the SKU in AMD Software (Adrenalin) or Device Manager. The HIP SDK Windows installer can bundle a display driver — we do not pin an Adrenalin version. Install the LunarG Vulkan SDK that llama.cpp’s Windows Vulkan docs require, then run vulkaninfo and look for the discrete Radeon. If vulkaninfo is empty, stop; do not guess a driver. BIOS notes (Above 4G, ReBAR/SAM, CSM off) apply on the same board. Cited, not measured on the local desktop.
Consumer first token on Windows: llama.cpp with Vulkan. Official build.md documents GGML_VULKAN=ON after the Vulkan SDK. Releases publish a Windows x64 Vulkan build. HIP SDK is cited for cards AMD lists, including R9700 — use it only after Vulkan tokens work. Do not start with Ollama’s bundled ROCm on gfx1201; a cited Windows report filtered that device to CPU while Vulkan still generated. Strata and vLLM are not a Windows first-run on AMDai.
Download the same official-instruct GGUF the Linux recipes name. Start at Q6, Q5, or Q4 and check VRAM + KV on the calculator. Q3 stays a cited labeled estimate unless a Windows run card exists — none does in this repo. Q1/Q2 stay off. Tok/s, VRAM, and RAM on /benchmarks and /results are Ubuntu llama-bench on the local desktop. Treat them as comparison, not a Windows promise.
The calculator prints Minimum system RAM needed (the load at the context you picked) and RAM needed at full context (262144 for the house 35B). Overnight UD-Q5 one-GPU Ubuntu peaks are 37.32 GB at 4K and 42.9 GB at 256K — measured on the local desktop, not a Windows kit. Keep the GGUF on NVMe.
Run official instruct in llama.cpp Vulkan. Confirm Task Manager or GPU-Z shows the Radeon busy. We do not publish a Windows launch line or a Windows tok/s. After tokens flow, read the Ubuntu cards so you know what “good” looked like on the measured path. A harness comes after loopback /v1 — that attach is also unmeasured on Windows here.
Radeon Stack Lab is the Ubuntu companion that records amdai-bench/run@1 cards. There is no public download URL and no Windows package. Open in Stack Lab uses the allowlisted amdai-stack-lab: scheme. Use it from a Linux box, or dual-boot, if you want house cards. Do not invent a Windows Stack Lab build.
Start from a Windows symptom on /troubleshoot (Windows ROCm filtered to CPU, Ollama gfx1201). Paste a log in the error matcher. Cited rows also live on /rocm-problems/windows-wsl-preview. Firmware: /bios. Jargon: /glossary. Short answers: /faq.
Use the calculator before you download a second quant. The two labels stay
Minimum system RAM needed
and
RAM needed at full context.
The house 35B peaks below are Ubuntu, not Windows.
If the app is not installed, packages are not published yet. There is no public download URL. Read the
Stack Lab install notes
.
If Stack Lab is already running on this machine,
open the local app
.