AMD funded a drop-in CUDA implementation built on ROCm: It's now open-source (phoronix.com)

Article summary

AMD funded a project to create a drop-in CUDA implementation built on ROCm, allowing many NVIDIA CUDA applications to run on AMD GPUs without modification. The project, called ZLUDA, has been made open-source after AMD discontinued funding. ZLUDA enables CUDA-enabled software to run on HIP/ROCm without any developer intervention. The implementation is not 100% fail-safe, but it is surprisingly capable for a single-developer effort.

Main themes

• AMD's CUDA implementation
• ROCm and GPU software development
• NVIDIA's API and ecosystem
• AMD's business strategy and management
• GPU compatibility and adoption
• Software development and optimization

What commenters say

• AMD's decision to discontinue funding for the ZLUDA project is puzzling, given its potential to increase adoption of AMD GPUs.
• The existence of a second implementation of CUDA, such as ZLUDA, could disincentivize NVIDIA from making breaking changes to the API.
• AMD's GPU division is underperforming due to poor management and a lack of focus on software development.
• The company's priorities are misplaced, with too much emphasis on ROCm and not enough on ensuring compatibility with existing software ecosystems.
• Having a CUDA wrapper on top of ROCm does not hinder ROCm's function as a GPGPU driver, and AMD should support both initiatives.
• AMD's hardware is competitive, but its software stack is lacking, making it difficult for the company to compete with NVIDIA.
• The ability to support a wide range of GPUs, including older models, is crucial for competing with NVIDIA, and AMD is falling short in this regard.