Andrzej Janik, a developer working on a tool that allowed Nvidia's CUDA code to run on AMD and Intel GPUs without any modifications, has open sourced his creation after support for the project was dropped by both AMD and Intel.
Although ZLUDA received a major update last week, bringing it to version 3, its future going forward is now in doubt.
ZLUDA originally emerged in 2020 as a promising tool to enable Intel GPUs to run CUDA, the proprietary hardware-software ecosystem that underpins Nvidia's dominance. Janik lobbied for Intel to adopt it, but the company, who he worked for at the time, decided there was no business case for running CUDA applications on its GPUs and requested the project be taken down.
AMD picks up the baton
Janik subsequently left Intel and was contracted by AMD to continue ZLUDA development. However, AMD also concluded there was no business case for running CUDA applications on its GPUs and ended its support for the project. Janik was released from the contract and was able to bring ZLUDA back to the public domain.
The current version of ZLUDA is significantly different from its 2020 iteration. Instead of being built on Intel's oneAPI and supporting Intel's GPUs, it is now based on AMD's ROCm solution and only supports Radeon GPUs. Janik stated that the project is more or less complete and will only receive updates for workloads he is personally interested in.
The fact that neither Intel nor AMD are interested in making their GPUs compatible with the existing CUDA ecosystem is significant. Both companies appear to prefer competing directly with CUDA using their own open-source solutions, oneAPI and ROCm, despite CUDA's continued popularity in professional and datacenter graphics software.
While AMD had been quietly funding ZLUDA for the past two years, the company decided to discontinue its support this year for unknown reasons. It’s possible that AMD wanted to avoid any possible lawsuits and so pulled out once the contract ended, meaning it couldn’t be directly tied to the project.
Despite the lack of corporate backing, ZLUDA has shown promise in testing, with many CUDA software able to run on HIP/ROCm without any modifications. However, as Phoronix points out, it's not a fail-safe solution, with some features such as NVIDIA OptiX not being fully supported.
