PostHeaderSection

Symbol

Global NavBar

PostContentSection

Global Footer

Blog Detail Page

Alex Medick

Discover how ZLUDA enables CUDA apps to run on AMD & Intel GPUs—no rewrites required. Learn the origin, challenges, and what’s next for ML. 

How CUDA Can Run on AMD & Intel GPUs

CUDA on non-NVIDIA GPUs isn’t sci-fi anymore. 

In this Beyond CUDA episode, Jeff Tatarchuk (TensorWave) chats with Andrzej Janik (creator of ZLUDA), a compatibility layer that runs many CUDA apps on AMD and Intel GPUs with no code changes. We dig into the origin story (hello-world moment in 2020), the Intel → AMD chapter, the open-source drama, and where the project is headed next: real ML workloads on desktop AMD, with an open invite for contributors.









Summary

Andrzej didn’t start as a GPU wunderkind. He came up through compiler QA at Intel, got drafted into GPU optimization, and fell in love with the low-level weeds. During lockdown, curiosity + Intel’s Level Zero API sparked a weekend project: implement enough CUDA to run on Intel GPUs. He expected it to fail. It didn’t. By late 2020 he had a green “hello world,” then a closed-source commercial app running at respectable performance.

That proof lit a fuse. ZLUDA on AMD GPUs began enabling professional creator apps (photogrammetry, etc.). The heavy boss battle was Arnold (renderer) with OptiX-style ray-tracing semantics - messy, complex, but Andre got real scenes rendering (with a few pixel quirks). Then the plot twist: priorities shifted; funding stopped; code briefly went open, then lawyers got involved; Andre rolled it back.

Today, ZLUDA has new backing and a focused roadmap: ML frameworks (PyTorch/JAX/TensorFlow) on AMD desktop GPUs. The big blocker isn’t the concept—it’s libraries and runtimes. To unlock broad, fast compatibility, the ecosystem needs cuDNN-parity APIs, stronger BLAS/DNN paths, and sturdier runtime features (e.g., textures). Andre’s message to the community: help is welcome—whether you care about Blender, datacenter GPUs, or even restoring particle effects in older games. Find him on Discord and GitHub.


Key Takeaways

 * Viability proven: You can run many CUDA workloads on non-NVIDIA GPUs. The idea isn’t theoretical anymore.
 * Libraries are king: Performance and breadth hinge on DNN/BLAS libraries and runtime maturity—not just kernel translation.
 * Strategy vs. UX: Vendor incentives don’t always align with user portability. Open layers like ZLUDA increase choice.
 * Near-term focus: ZLUDA’s 2025 push is ML on AMD desktop GPUs with incremental compatibility/perf gains for PyTorch/JAX/TF.
 * Call for contributors: Want CUDA portability for your stack (rendering, ML, games)? Join the Discord/GitHub and plug in.
 * Why it matters for teams: Portability de-risks supply, improves leverage on cost/perf, and accelerates multi-vendor strategies without rewrites.
 * For builders on AMD: If you’re exploring an AMD-first path, projects like ZLUDA plus robust libraries are the unlock for “drop-in” workflows.

Build on AMD today: Talk to us about running your CUDA-based stack on AMD Instinct™ GPUs with zero code changes.


About TensorWave

TensorWave is the AI neocloud purpose-built for performance. Powered exclusively by AMD Instinct™ Series GPUs, we deliver high-bandwidth, memory-optimized infrastructure that scales with your most demanding models—training or inference.

Ready to get started? Connect with a Sales Engineer.

How CUDA Can Run on AMD & Intel GPUs

Product

Solutions

Resources

Company

Product

Resources

Solutions

Company