New ZLUDA-based stack runs CUDA workloads on AMD GPUs under Windows

A newly published GitHub project stitches together ZLUDA, AMD’s HIP/ROCm SDK, and LibTorch into a reproducible Windows setup that lets CUDA-targeted compute applications run on AMD hardware. The stack translates CUDA library calls (cuBLAS, cuBLASLt, cuSPARSE, cuFFT) to their ROCm equivalents, and an install script pins specific versions — ZLUDA v6-preview.69, HIP SDK 6.4, LibTorch 2.3.0+cu118 — downloads the components, verifies SHA-256 hashes, and runs ZLUDA’s own compatibility check against the installed AMD stack. The maintainer reports a real proof point: a 2.2-million-parameter PPO reinforcement-learning network completed forward passes, learning, and optimizer steps for 65,536 timesteps on the AMD device.

The important caveat is scope. Only a single card, the Radeon RX 9060 XT (gfx1200, RDNA4), is validated; every other GPU is flagged as an untested candidate rather than a supported device. Because Windows exposes just a subset of ROCm, the stable HIP SDK ships without MIOpen/cuDNN, so convolution-heavy models are likely to break, and NCCL, TensorRT, and unusual PTX behavior may fail outright. Coverage is fundamentally workload-dependent — dense GEMM-style training works, but this is not a drop-in CUDA replacement.

Notably, the project deliberately sticks to an upstream-only path and declines to ship a set of “recovered” custom DLLs, citing incomplete provenance and bundled third-party AMD binaries. A controlled A/B benchmark backs that choice on the merits too: the clean upstream runtime hit a median 13,278 steps per second versus 12,876 for the custom overlay, making the overlay both legally murkier and about 3% slower. It’s a pragmatic, honestly-scoped attempt at loosening NVIDIA’s CUDA lock-in on consumer AMD silicon, though the narrow hardware validation keeps it firmly in early-adopter territory.