Andrzej Janik, a developer engaged on a device that allowed Nvidia’s CUDA code to run on AMD and Intel GPUs with none modifications, has open sourced his creation after help for the mission was dropped by each AMD and Intel.
Though ZLUDA acquired a serious replace final week, bringing it to model 3, its future going ahead is now unsure.
ZLUDA initially emerged in 2020 as a promising device to allow Intel GPUs to run CUDA, the proprietary hardware-software ecosystem that underpins Nvidia’s dominance. Janik lobbied for Intel to undertake it, however the firm, who he labored for on the time, determined there was no enterprise case for working CUDA functions on its GPUs and requested the mission be taken down.
AMD picks up the baton
Janik subsequently left Intel and was contracted by AMD to proceed ZLUDA growth. Nevertheless, AMD additionally concluded there was no enterprise case for working CUDA functions on its GPUs and ended its help for the mission. Janik was launched from the contract and was in a position to carry ZLUDA again to the general public area.
The present model of ZLUDA is considerably completely different from its 2020 iteration. As an alternative of being constructed on Intel’s oneAPI and supporting Intel’s GPUs, it’s now primarily based on AMD’s ROCm answer and solely helps Radeon GPUs. Janik acknowledged that the mission is kind of full and can solely obtain updates for workloads he’s personally taken with.
The truth that neither Intel nor AMD are taken with making their GPUs appropriate with the prevailing CUDA ecosystem is important. Each firms seem to favor competing instantly with CUDA utilizing their very own open-source options, oneAPI and ROCm, regardless of CUDA’s continued reputation in skilled and datacenter graphics software program.
Whereas AMD had been quietly funding ZLUDA for the previous two years, the corporate determined to discontinue its help this 12 months for unknown causes. It’s attainable that AMD wished to keep away from any attainable lawsuits and so pulled out as soon as the contract ended, which means it couldn’t be instantly tied to the mission.
Are you a professional? Subscribe to our e-newsletter
Signal as much as the TechRadar Professional e-newsletter to get all the highest information, opinion, options and steerage your corporation must succeed!
Regardless of the shortage of company backing, ZLUDA has proven promise in testing, with many CUDA software program in a position to run on HIP/ROCm with none modifications. Nevertheless, as Phoronix factors out, it isn’t a fail-safe answer, with some options corresponding to NVIDIA OptiX not being absolutely supported.
<header
