AMD funded a drop-in CUDA implementation built on ROCm: It's now open-source

Original Article ↗ Hacker News Discussion ↗

AMD quietly funded ZLUDA, a drop‑in CUDA implementation on top of its ROCm stack that can run many NVIDIA‑targeted workloads on Radeon GPUs, in some cases even outperforming AMD’s own native HIP backend. The project has now been open‑sourced after AMD ended its contract, sparking debate over whether abandoning a near‑ready CUDA compatibility layer is a strategic blunder at a time when NVIDIA dominates AI and GPU compute. Commenters contrast NVIDIA’s strong, widely supported software ecosystem with AMD’s fragmented ROCm support and argue over whether AMD should chase CUDA compatibility, double down on its own stack, or focus on high‑end data‑center deployments instead of consumer GPUs.

Project & Release

  • ZLUDA is a “drop‑in” CUDA implementation on top of AMD’s ROCm/HIP stack, originally funded by AMD under a private contract.
  • AMD canceled funding after ~2 years, stating there was “no business case” for running CUDA on AMD GPUs; contract allowed the author to open‑source the code, resulting in a large one‑shot commit.

Performance & Capabilities

  • Benchmarks cited in the thread show Blender’s CUDA backend via ZLUDA sometimes outperforming Blender’s native HIP backend on Radeon hardware.
  • Current cuDNN coverage is minimal (enough to run ResNet‑50); PyTorch support is described as very limited.
  • ZLUDA re‑implements large parts of the CUDA “dark API” via reverse‑engineering, which is complex and brittle by nature.

Reactions to AMD Ending Funding

  • Many commenters think dropping the project is strategically irrational given NVIDIA’s AI boom and CUDA lock‑in.
  • Others argue AMD doesn’t want to be permanently chasing a proprietary API they don’t control and prefers to push ROCm/HIP directly.
  • Some suspect internal politics, resource constraints, or legal‑risk aversion; others note both Intel and AMD have previously concluded “no business case” for CUDA compatibility.

State of ROCm and AMD GPU Strategy

  • Heavy criticism that ROCm officially supports very few consumer GPUs, is hard to install, and has stability issues; this deters developers and hobbyists.
  • Defenders note ROCm unofficially works on many more cards, and AMD is prioritizing data‑center/HPC (MI300, Top500 systems) over gaming/enthusiast compute.
  • Multiple anecdotes of labs and individuals choosing NVIDIA cards solely due to CUDA and tooling, despite liking AMD hardware.

CUDA’s Moat and Alternatives

  • Broad consensus that CUDA’s dominance comes from years of investment in compilers, libraries (cuDNN, cuBLAS), tools, and backward compatibility.
  • Debate whether a compatibility layer strengthens CUDA’s position (like OS/2 running Windows apps) or is a necessary bridge that could later enable “embrace/extend”.
  • Alternatives mentioned: ROCm/HIP, SYCL, Vulkan compute, WebGPU, OneAPI; OpenCL is widely seen as a failed attempt.

Open‑Source Prospects & Unclear Points

  • Optimism that open‑sourcing ZLUDA lets the community continue development and pressures NVIDIA on API stability.
  • Skepticism about long‑term sustainability: huge maintenance burden, constantly changing CUDA, and limited AMD commitment.
  • Legal status of re‑implementing CUDA APIs/ABIs is debated; implications remain unclear.