Heterogeneous computing is clearly here to stay but now’s the time to get down to brass tacks and start addressing standards, portability, and other elements common to maturing technologies.
The need to integrate various types of processors, including CPUs, GPUs, and specialized chips for AI and other applications is clear, thus today we’re met with news about the inception of the Unified Acceleration (UXL) Foundation, as part of the broader Linux Foundation.
Its founders think the effort marks the beginning of a much-needed era of open standards, and multi-vendor, multi-architecture collaboration.
At the helm of this groundbreaking initiative is Rod Burns, Vice President of Ecosystem and Chair of the UXL Foundation Steering Committee. Burns, who has been heavily involved in the accelerator programming community through his tenure at Codeplay, described the growing necessity for a standards-based software stack, especially given the significant uptick in GPU usage facilitated by data-intensive workloads and Large Language Model (LLM) based AI applications.
The momentum GPUs have gained over the past decade is irrefutable, with a substantial representation in the world’s top supercomputers and of course in AI training and inference. However, the GPU software stack, particularly in AI domains, is in nascent stages with evolving standards.
It is at this juncture that the UXL Foundation emerges, as Burns says, with the goal of fostering performance portability and a multi-vendor programming model built upon open standards and source projects.
Hosted by the Linux Foundation’s Joint Development Foundation, the UXL Foundation integrates industry stalwarts like Arm, Fujitsu, Google, Imagination Technologies, Intel, Qualcomm Technologies Inc, and Samsung with more likely to come.
The platform aims to build upon the strides made through SYCL, a Khronos open standard, and the oneAPI specification and projects, all of which target an open standard programming for accelerators. The synergy between oneAPI and SYCL, especially facilitated through the DPC++ compiler based on LLVM Clang, forms a crucial aspect of this venture, enabling a cohesive and power-optimized programming model, as Burns describes in his kickoff statement.
The UXL Foundation is setting about to outline clear objectives and strategies, centered around vendor neutrality and independence. The foundation imagines a cohesive development platform for heterogeneous architectures, with collaboration within the community to evolve open-source implementations and meet universal needs.
A focal point of the foundation’s approach is the facilitation of Special Interest Groups (SIGs) that focus on various domains like AI, Hardware, Image, Language, and Math. These groups aim to drive the initial transitions of open-source projects to the UXL Foundation, articulating and executing goals and roadmaps throughout 2023.
Launched officially on September 19, 2023, at the Linux Foundation Open Source Summit held in Bilbao, Spain, the UXL Foundation does have its work cut out.
The mission of the UXL Foundation is twofold: to construct a multi-vendor software ecosystem for all accelerators and to unify the heterogeneous computing sphere through open standards. This initiative is a natural evolution of the oneAPI project, which was formed to simplify the development of performant, cross-platform applications through an open standard accelerator programming model.
This new foundation operates through Working Groups and Special Interest Groups (SIGs), fostering collaboration and innovation in the space of accelerated computing. The strategic formation of the UXL Foundation marks a pivotal step in leveraging the oneAPI specification across the industry, bringing together a consortium of industry leaders to cultivate a cross-architecture unified programming model.
The involvement of figures like Robert Cohn, the oneAPI Specification Editor from Intel Corporation, echoes the sentiment that open source and standards are critical in developing a cross-platform software stack for GPUs and other accelerators. This perspective aligns with the transformation Linux and GNU spearheaded in the CPU software stack, aiming to lay the foundation for the upcoming generation of data and computationally intensive applications.
The foundation seeks to not only define a vendor-neutral development platform but also to foster a community that works synergistically to meet universal needs through open-source implementations.
For those keen on being part of this transformative journey, the UXL Foundation welcomes participation through its various Special Interest Groups and Working Groups. You can find a starting point at UXLFoundation.org.
To Cure Iron Anemia With SYCL, Intel Buys Codeplay
Intel doesn’t want to just create a rival to the CUDA programming model and library stack so it can better compete against Nvidia in the GPU compute market. With oneAPI, it wants to create an open ecosystem that includes a programming framework, called oneAPI and largely based on the SYCL …
Intel Takes The SYCL To Nvidia’s CUDA With Migration Tool
After years of false starts and delays with various products, we are finally at a point where Intel will truly start to test the breadth of its heterogenous computing strategy, thanks to the release of new Gaudi2 machine learning chips from Intel and the upcoming launch of its much-anticipated “Ponte …
Can SYCL Slice into Broader Supercomputing?
There are a few unignorable trends in high performance computing, especially in the exascale age. First, heterogenous and differing architectures at the major HPC sites are diverging with some using AMD CPUs and GPUs, others Nvidia, and of course, still others sticking with the well-tread Intel path. Second, codes are …
The UXL foundation’s mission seems like a laudable goal, a bit like CXL as discussed in the “Iron Anemia” TNP article (06/01/22), but with Codeplay and SYCL as protagonists. On the other hand, this being software, the incentive to adopt a Connor MacLeod highlander perspective on converged standardization is probably not as strong as it is for hardware interoperability.
Seeing the success of kokkos in SCREAM, and the aspirations of Mojo (“in the end, there can be only one”), suggests that UXL could benefit by their support (LLVM seems to be there with DPC++). Words of support from folks involved in the Exascale Computing Project (ECP) wouldn’t hurt either I think (not to mention IBM, AMD, Cerebras, SambaNova, …).
In my view, beyond a unified framework and the Linux Foundation, seducing coders away from CUDA is probably going to require energetic gamer-style digestive enthusiasm (ATI Rage!), an atmosphere of fun, some look-we-can-do-anything!-we’re-on-top-of-the-world! attitude, and a massive flood of easy-to-follow online tutorials showcasing 1000x speedups over naive unoptimized code.