-
The Linux Foundation launched the new Unified Acceleration Foundation to tackle CPU, GPU and FPGA programming complexity with an open source spec
-
Founding members include Google Cloud, Arm, Intel and more
-
The group wants to recruit more members, especially cloud providers
Heavy hitters including Arm, Google Cloud, Intel, Qualcomm, Samsung, Fujitsu and Imagination Technologies banded together under the guise of the new Unified Acceleration (UXL) Foundation with the goal of making life just a little easier for developers working with artificial intelligence (AI) and other high-performance workloads.
The UXL Foundation was formed under the umbrella of the Linux Foundation. The group’s primary aim is to create a single, open-source standard for accelerator programming that spans CPUs, GPUs and FPGAs from different vendors.
“Ultimately, the goal of this foundation is to get things in such a place that software developers can focus on developing software and not worry about the hardware they’re running on,” Rod Burns, UXL Foundation Steering Committee Chair and VP of Ecosystem at Codeplay Software, told Silverlinings.
Burns noted the hardware question has become a major issue as the industry – and certain verticals like healthcare, finance and research – has moved toward heterogenous computing, which involves computers containing multiple different kinds of processors. The challenge is there’s no specification that spans all of these processors today. Thus, a developer can write software for an accelerator or GPU from one vendor, but if they want move to a different accelerator they have to rewrite their code.
He added that the U.S. National Laboratories is one entity which has encountered this issue. While the supercomputer at one lab location runs on Intel CPUs and GPUs, another runs on NVIDIA and still another on Arm GPUs. That makes it harder for researchers to pool resources and deploy software across those different machines.
With the rise of AI driving demand for GPU computing, Burns said the issue is also a “challenge” for cloud providers.
“They want to be able to provide their GPU-based architectures to developers who may not necessarily want to use the same GPU architecture all the time. They many want to deploy specialist accelerators in the cloud. So I think there is certainly a lot of interest in a unified approach to that,” he said.
Google Cloud is already on board but Burns noted the group is hoping to recruit more vendor participation, particularly from cloud service providers.
What’s next
The group is starting off with an existing set of assets contributed by Intel, including its oneAPI spec and the open source code used to implement that. In the short term, Burns said the focus will be on the “smooth migration of the open-source code bases to the foundation and the spec alongside that.”
The foundation’s steering committee will also be having a look at what gaps there may be in the Intel spec and what new features need to be added. From there, it can set priorities and goals and establish working groups.
In the medium term, Burns said the foundation will also work to get even more vendors onboard. It’s not yet clear when the group’s first release might be published.