CUDA-beating ZLUDA breathes new life with financial backing from unknown party — pivots to AI workloads across multiple GPU vendors | usagoldmines.com

ZLUDA, an open-source CUDA translation layer, has lived two fairly vivid lives with Intel after which AMD GPUs. It was practically killed in August when AMD requested to take down the code developed utilizing its funds. Nonetheless, as its developer, Andrzej Janik, secured funding from a mysterious sponsor, ZLUDA now has a 3rd life. This time round, the main target of ZLUDA might be to run AI/ML software program designed for CUDA GPUs on processors from different distributors utilizing a translation layer, stories Phoronix.

ZLUDA was initially designed to run inventive skilled CUDA-based purposes on Intel after which AMD GPUs, whereas the upcoming iteration of ZLUDA shifts focus to accommodate AI and machine studying workloads. Additionally, the emphasis is not simply on Intel or AMD. As an alternative, it presents a number of GPU vendor assist, making ZLUDA relevant throughout completely different GPU architectures. Nonetheless, in the meanwhile, most growth efforts are targeting AMD GPUs, significantly RDNA1 and newer architectures. Assist is being constructed round AMD’s ROCm 6.1+ compute stack, laying the muse for broader, multi-architecture compatibility sooner or later.

Andrzej Janik is at present working to make AI/ML frameworks like PyTorch, TensorFlow, and Llama.cpp operate seamlessly utilizing CUDA on non-Nvidia GPUs utilizing his translation layer, in line with Phoronix, who spoke to the developer. Janik predicts it is going to take a couple of 12 months to develop the brand new ZLUDA code to a steady state the place it could actually successfully deal with AI/ML workloads throughout a number of GPUs. Contributions from the open-source group might be welcomed because the mission evolves. So, ZLUDA will stay open supply, or no less than it seems so at the moment.

Though ZLUDA now has a monetary backer, the sponsor has chosen to stay nameless for now. We are able to solely speculate who the sponsor is as a result of they should run AI workloads at scale and opted for multi-GPU vendor assist. Additionally, we presume it’s sufficiently big to not be afraid of getting right into a battle over working CUDA software program via a translation layer, which Nvidia doesn’t endorse nowadays. But, the developer says that this ‘stealth’ sponsor is anticipated to be revealed later, offering extra perception into the path and future assist of ZLUDA.