CUDA on AMD, on Windows: ZLUDA v6 now runs llama.cpp near-native
A trending Windows recipe pairs ZLUDA v6 with ROCm to run CUDA-only apps on a Radeon; inference flies, but cuDNN gaps still block most training.
Copy markdown
CUDA-only apps, now on a Radeon
A GitHub recipe trending on Hacker News wires ZLUDA v6 to AMD's HIP/ROCm so unmodified CUDA Windows apps run on a Radeon — no NVIDIA card, no code changes. Phoronix clocks llama.cpp at near-native ROCm speed and calls the Windows build finally “acceptable quality.”
The proof: 13,278 steps/sec on an RX 9060 XT
A Sept 13 benchmark ran a PyTorch/LibTorch PPO network on an RX 9060 XT at 13,278 median steps/sec via the upstream ZLUDA v6-preview.69 + HIP SDK 6.4 path, with cuBLAS, cuSPARSE and cuFFT all live. It's a reproducible recipe, not a vague “it works.”
The catch: no cuDNN, one GPU, one maintainer
cuDNN still isn't available on the Windows HIP path, which rules out most standard neural nets and training; NCCL, TensorRT and custom PTX kernels are out too. Only the RX 9060 XT is validated, and ZLUDA is unfunded again — “back to a weekend project,” per creator Andrzej Janik. Linux ROCm stays far ahead.
Build this weekend: local inference on the card you own
Got a Radeon and locked out of CUDA-only tooling? This is the unlock — run llama.cpp or ComfyUI/Stable Diffusion locally without buying NVIDIA. Treat it as an inference box, not a training rig, and pin the exact ZLUDA and HIP SDK versions from the recipe.
Elsewhere: Docket logs what your coding agent actually did
A new open-source tool, Docket, writes a per-commit evidence record for agent-written code — what the agent tried, what verified it, and what nobody reviewed. Early (v0.0.4) but aimed at the “who checked this AI PR?” problem as agent commits pile up.