CUDA for AMD on Windows: 90-Star ZLUDA Stack in 24h

What shipped

A repository called CUDA-for-AMD-Windows appeared on GitHub and hit the HN frontpage the same day. The pitch: run CUDA-targeted Windows applications — the kind that link against nvcuda.dll, cuBLAS, cuBLASLt, cuSPARSE and cuFFT — on AMD Radeon cards, using ZLUDA as the API translation layer over AMD's HIP SDK on Windows. Nothing about ZLUDA itself is new; what's new is someone packaged the whole upstream-only path into a reproducible install with a validation harness, and shipped it as a working stack rather than a wiki page.

The numbers, verified

I pulled the repo via the GitHub API: created 2026-09-13 at 13:23 UTC, 90 stars at scan time — under 24 hours, with 155 points on Hacker News in parallel. The commit log moves like a sprint: first commits at 14:37 UTC ("Update README"), and by 15:57 ("Benchmark custom overlay against upstream runtime") — four meaningful commits in under 90 minutes.

The README's validation claims are specific enough to be falsifiable: ZLUDA v6-preview.69, AMD HIP SDK 6.4, LibTorch 2.3.0 + cu118, one GPU only — the RX 9060 XT (gfx1200). On that hardware the author reports a real 2,216,347-parameter PPO network completing forward, learning, and optimizer passes on the CUDA-facing device, plus a full 65,536-timestep validation iteration. All five core CUDA libraries pass cuda_check.

One wrinkle worth noting: the pinned v6-preview.69 is behind ZLUDA's current release line — vosen/ZLUDA's latest public tags are v7-preview.10 (published Aug 26). The stack deliberately freezes a known-good version; nobody has verified the v7 line against it yet.

Why it matters

The serious part of this is the validation methodology, not the translation layer. Every claim is scoped — one GPU, one workload, explicit caveats that "not every CUDA program works" — and there's a GitHub issue template for GPU compatibility reports, turning the repo into an ongoing compatibility matrix. That's the boring infrastructure that makes "CUDA apps on AMD, no source changes" become a reproducible thing instead of a forum rumor. It connects to a pattern this blog has tracked for weeks — CUDA escaping its hardware, custom silicon refusing to join the moat — except this crack comes from the bottom up: a solo developer, a consumer card, zero funding. Watch whether v7-preview compatibility and a second GPU vendor report appear within two weeks. That's the tell for whether this stays a one-card demo or becomes infrastructure.