By ByteBot
Andrzej Janik released ZLUDA v6 on June 29. The update is genuinely impressive: 32-bit PhysX support that triples frame rates in legacy games on AMD hardware, matrix multiply-accumulate instructions for machine learning, ROCm 7 compatibility, and fixes for PyTorch, llama.cpp, and vLLM. It is also the moment Janik announced that commercial funding for ZLUDA has ended — again — and the project is back to being a weekend hobby. That is the third time this has happened.
What ZLUDA Does — and Why It Keeps Attracting Money
ZLUDA is a drop-in CUDA compatibility layer. It intercepts CUDA API calls and redirects them to AMD’s ROCm platform, letting unmodified CUDA binaries run on AMD GPUs without recompilation. That is a fundamentally different proposition from AMD’s official HIPIFY tool, which requires developers to port their code. ZLUDA’s pitch is zero friction: take your CUDA workload, run it on AMD hardware, pay less.
The appeal is obvious to anyone watching AI infrastructure costs. Nvidia holds roughly 85% of the AI GPU market as of mid-2026 — and that dominance is not purely about hardware performance. It is about the CUDA ecosystem: 20 years of optimized libraries, millions of developers writing CUDA-native code, frameworks like PyTorch built CUDA-first. Even as AMD’s Instinct MI300X and MI350X chips have become genuinely competitive on raw compute, the software gap keeps enterprise buyers locked to Nvidia. ZLUDA offered a shortcut through that gap.
The Pattern Nobody Can Break
This is the third time a funding source has concluded that ZLUDA is not worth sustaining. In 2020, Janik built the original version at Intel to target Intel GPUs. Intel shut it down. AMD hired Janik in 2022 and funded a rebuild for AMD hardware — then quietly stopped paying in early 2024. When Janik tried to release the AMD-funded codebase publicly, AMD’s legal team demanded it be taken down, claiming rights to code Janik believed his contract allowed him to release. Janik’s response was pragmatic: “the legality of emails is unimportant. The choice is between a rewrite — cheaper, guaranteed result — and possibly fighting it in a court.” He rewrote it.
In late 2024, an undisclosed sponsor — widely suspected to be an AI company that wanted to run CUDA workloads on AMD Instinct at lower cost — funded a second commercial push. That money is now gone too. A second developer on the project, Violet, is now listed as “Developer Emeritus.” The structural problem is that nobody with sustained funding has an incentive to keep this alive. Nvidia has every reason to let the CUDA moat stand. AMD would prefer developers use ROCm natively rather than depend on an unofficial shim. Commercial AI companies, it seems, will fund a CUDA workaround exactly as long as it is cheaper than switching properly, and not a day longer.
What v6 Actually Delivers
Despite the funding news, v6 is a solid release. The headline feature is 32-bit PhysX support, which enables GPU-accelerated physics in older titles that previously required Nvidia hardware. Mafia II, running on an AMD RX 9070 XT, jumped from 26 FPS to 80 FPS with PhysX enabled — a 3x improvement. The support is pre-alpha and fluid simulations remain glitchy, but it works well enough to be useful. For AI workloads, the release adds matrix multiply-accumulate (MMA) instructions with fallback support for RDNA1 and RDNA2 architectures, MIOpen integration, expanded cuBLAS compatibility, and fixes for PyTorch, llama.cpp, vLLM, and KataGo. Full release notes are on GitHub.
What Comes Next for ZLUDA
Janik has been explicit about what “weekend project” means in practice: development continues at a pace he can sustain, updates will come less frequently, and priorities will follow his personal interest rather than commercial demand. The project blog remains active and the codebase is open source under the MIT license, so the community can contribute.
Whether ZLUDA ever achieves sustained backing depends on whether the broader CUDA alternative problem gets solved at a higher level. AMD’s ROCm is genuinely improving — analysts called it “actually good” in 2025 after years of derision. The inference workload shift toward less custom kernel dependency may reduce CUDA’s structural advantage over time. For now, however, the CUDA moat is intact. Janik shipped ZLUDA 6 anyway. After everything AMD put him through, that persistence deserves acknowledgment.
