2023 marked a watershed for the oneAPI industry initiative with the Linux Foundation’s announcement of the Unified Acceleration (UXL) Foundation—“a cross-industry group delivering an open standard accelerator programming model”—adopting the oneAPI specification as its starting point.
At the December 2023 oneAPI DevSummit for AI and HPC, Intel’s Joe Curley and Sanjiv Shah delivered an inspiring retrospective of this and other accomplishments that shaped the past year for the oneAPI community and set the direction for an exciting year ahead.
Shah, VP of Intel’s Software and Advanced Technology Group and General Manager of Developer Software Engineering, opened with the observation that for the past four decades, developers have been able to run computing languages on any device, from PCs and embedded devices to mainframes and supercomputers. The introduction of accelerator technologies—GPUs, FPGAs, and specialized ASICs for AI—broke the model of code portability by forcing developers to write architecture-specific code that can only run on the device it was written for. “Why should the world of accelerators be any different?”
Developers want to take advantage of the unprecedented capabilities offered by accelerators, but their dilemma has been having their choice of accelerator restricted by proprietary software or having to bear the high cost of rewriting code for different target devices. The mission of the oneAPI programming model is to make accelerator programming open and portable.
“We started oneAPI with a vision of making accelerated programming portable and open, offering the benefits of accelerators to everyone.” oneAPI allows for a single codebase to be deployed across multiple computing architectures—CPU, GPU, FPGA, and others—through direct programming in C++ with SYCL,* together with a set of libraries targeting common functions like math, threading, and neural networks, and a hardware abstraction layer that enables programming in one language and target different accelerators. In addition, many popular AI frameworks including PyTorch* and TensorFlow* have built in oneAPI optimizations. All built in the open, through open source, open specifications, and open licenses. This provides developers performance, productivity, and freedom to program on the device of their choosing.
While Intel offers toolkits that optimize oneAPI for Intel products, one of the major developments over the past year was the introduction of plug-ins for Nvidia and AMD GPUs by Codeplay, which enable developers using Intel tools to target multiple vendor devices “so that we can truly live up to the portability promise.”
Together with proven portability, oneAPI offers proven performance. Shah provided a number of examples of applications run with SYCL on Nvidia and AMD GPUs that demonstrate comparable performance across a wide range of workloads compared to the vendor-native CUDA and HIP languages. In addition, developers can migrate CUDA code to SYCL through automated tools such as the Intel® DPC++ Compatibility Tool or open source SYCLomatic, enabling the original code to run on a variety of hardware. Over the past year, the growing oneAPI ecosystem has been building on oneAPI to support new languages and hardware beyond Intel, Nvidia, and AMD offerings.
Since the inception of oneAPI, Intel was “very clear upfront that…we would make this openly governed by neutral bodies that anybody in the world could join.” This promise has been fulfilled with the announcement in September of the Unified Acceleration (UXL) Foundation, a joint foundation project under the Linux Foundation, which now governs the oneAPI specification. Steering members include ARM, Fujitsu, Google Cloud, Imagination Technologies, Intel, Qualcomm, Samsung, and VMware. “The big benefit of working with the Linux Foundation,” said Shah, “is that they have a very fantastic, structured way of engaging with everybody…They have a well-defined governance structure where it really is a neutral body that everybody can work with.”
“We couldn’t be more proud to have brought oneAPI into the world and to be handing over all governance and the open source projects and the specification and how this project runs forward to UXL.”
Joe Curley, VP of Intel’s Software and Advanced Technology Group and General Manager of Software Products & Ecosystem, put the growing adoption of oneAPI in the context of Geoffrey Moore’s famous diagram of how new technologies get adopted in the marketplace from Crossing the Chasm. “It happens in phases. The first phase you have to identify a compelling vision.” In the case of oneAPI, it was the value to developers of being able to harness different kinds of accelerators for AI and other tasks without the costs and complexity of individual software stacks for each individual device. “So you’ve first got to first try to capture innovators who align to what you’re trying to do philosophically. And you work together with them to prove out that the idea is technically possible. That leads to a second phase of developer who shares the problem and now seeing proof of existence is willing to jump in and invest.”
The challenge is that these innovators and early adopters willing to collaborate on a new technology represent a small portion of the potential user base, while the majority of users will expect a finished product. Technologies take time to mature; crossing the chasm requires building up the technology so that the majority can see the value and eventually trust and adopt it in a broad way.
From that perspective, continued Curley, “we were really gifted by being able to a number of leaders in the marketplace” as exemplified by leading universities and research organizations that created oneAPI Centers of Excellence. “All of these people saw the need for multi-standard, multi-architecture programming of accelerators and were willing to join us early on.”
Curley pointed to a number of complex HPC applications built on oneAPI by these innovators—some showcased at Supercomputing 23—creating early success stories that compelled a second wave of early adopters to join in.
In particular, Curley noted the growing number of communities migrating codes to oneAPI with SYCL, and the fact that they’ve been able to port applications with ease and obtain comparable performance.
“The work that we've tried to do in SYCLomatic and…with our own subsidiary Codeplay is develop quality implementations that will run on other hardware. And the early partners have found that.”
This growing application base is sparking a larger wave of adoption, including ISVs and AI frameworks like TensorFlow and PyTorch deploying oneAPI and tools utilizing oneAPI to create scalable performance on CPUs and accelerators.
“So we're starting to see the movement where mass production and use of the oneAPI technologies has already started to occur and that's exceptionally rewarding as we've gone forward into our fourth year.”
To make oneAPI even more accessible, Intel announced general availability of the Intel® Developer Cloud, where developers can access the latest Intel hardware along with the latest software including Intel® Software Development Tools 2024.0—a complete set of advanced compilers, libraries, analysis and debug tools, and optimized frameworks built on oneAPI that simplify the development and deployment of accelerated computing solutions.
Ready to hop on board? Check out the links below to explore oneAPI tools and engage with the community!
CTAs:
View sessions and keynote from oneAPI DevSummit for AI and HPC
Get a quick intro to oneAPI in this video
Explore the latest Intel® Software Development Tools
Get started on Intel® Developer Cloud
Join the Unified Acceleration (UXL) Foundation
Become a oneAPI Innovator
Become a Student Ambassador
Apply to Intel Liftoff for Startups
*Other names and brands may be claimed as the property of others. SYCL is a trademark of the Khronos Group Inc.
Technology marketing leader with a passion for connecting customers to Intel value!