For AMD GPU owners running distributed computing workloads, the CUDA ecosystem has been a hard wall. Folding@Home's most optimized cores depend on NVIDIA's cuFFT library, and if your card doesn't speak that dialect, you're stuck on slower fallback paths. ZLUDA 7-preview.11, released 22 September 2026, cracks that wall for a specific slice: the FFT calls those CUDA cores make. AMD Radeon owners can now run the same optimized cores their NVIDIA counterparts use, without buying a second GPU.

The mapping is direct: where a CUDA application calls a cuFFT function, ZLUDA intercepts that call and redirects it to the corresponding hipFFT function in AMD's ROCm/HIP stack. Developer AmosKito1 handled the implementation. The pull request confirms the translation is one-to-one where the interfaces match. ZLUDA only covers cuFFT APIs that have a direct hipFFT equivalent; anything without a clean counterpart in AMD's library remains unimplemented. This is a translation layer, not a reimplementation. It works because hipFFT already exists as AMD's answer to cuFFT, and ZLUDA simply wires the two together so CUDA binaries don't need rewriting.

Before this release, Folding@Home's Core24 and Core27 — the two CUDA-optimized folding cores — were unavailable to AMD GPU users. The cores call cuFFT for their FFT operations, and without that library present, they fail to initialize. After ZLUDA 7-preview.11, those same two cores run on AMD hardware. The number that matters is two: Core24 and Core27. The source does not state whether other Folding@Home cores or other CUDA applications that use cuFFT benefit from this change. The consumer impact is concrete but narrow: if you run Folding@Home on an AMD card, you can now use the optimized cores instead of slower fallback paths.

The source does not provide performance numbers. There is no benchmark comparing hipFFT throughput against cuFFT on equivalent hardware, so the real-world speed of Folding@Home on AMD via ZLUDA remains unknown. The specific list of which cuFFT APIs are covered is not enumerated in the article; the pull request is referenced but its full contents are not reproduced. No AMD GPU model is named as tested or confirmed working. The release is labeled preview, which in ZLUDA's versioning suggests the implementation is not yet considered stable.

Watch for performance benchmarks from the Folding@Home community running Core24 and Core27 on AMD hardware through ZLUDA. If hipFFT matches cuFFT within a few percent, the practical gap between NVIDIA and AMD for this workload closes. If it trails significantly, the translation layer works but the speed penalty remains. The next ZLUDA release will likely expand cuFFT API coverage or add other NVIDIA library translations, but that is speculation beyond what the source states.

The cuFFT-to-hipFFT mapping is a thin translation layer that works only where interfaces are one-to-one, limiting its scope to a subset of FFT operations. Folding@Home's Core24 and Core27 are the confirmed beneficiaries; broader CUDA application compatibility remains untested.