Intel has expanded its effort to drive adoption of a set of open compilers, libraries, analysis and porting tools and frameworks for artificial intelligence (AI) applications, collectively made available via a oneAPI initiative, to include support for its latest generation of processors.
The oneAPI initiative was launched more than three years ago to create a set of multi-architecture tools for building applications that could be deployed on any CPUs, graphical processor units (GPUs) and field programmable gate arrays (FPGAs). Processors that now support oneAPI include forthcoming 4th Gen Intel Xeon Scalable processors, Intel Xeon CPU Max Series and Intel Data Center GPUs.
James Reinders, chief evangelist of Intel Software Products, said there is now a clear need for open tools and frameworks that provide equal access to multiple classes of processors. As part of that effort, the latest 2023 version of oneAPI, for example, adds support for plug-ins for Nvidia and AMD GPUs developed by Codeplay, a unit of Intel, to simplify writing SYCL code in a way that makes code portable across additional processor architectures.
There is also a CUDA-to-SYCL code migration capability that supports more than 100 CUDA application programming interfaces (APIs) using the Intel DPC++ Compatibility Tool based on an open source SYCLomatic project. Intel is also providing tools to identify imbalances at scale using Intel VTune Profiler, and an Intel Advisor tool now provides automated roofline analysis for Intel Data Center GPU Max Series to identify root causes of memory, cache or compute bottlenecks along with insights for optimizing data transfer costs involving CPU-to-GPU offloading.
As IT continues to evolve, it will become more routine for developers to invoke heterogeneous classes of processors within the same application, but it’s been challenging to convince providers of other types of processors to support the oneAPI initiative, noted Reinders. Processor providers have historically relied on compilers and libraries to lock developers into specific classes of processors.
However, it’s becoming more common for developers to build applications that invoke multiple types of processors. The Intel oneAPI initiative is intended to simplify access to heterogeneous processors, also known as XPUs. That’s critical, because the pace of processor innovation has accelerated to the point where developers need to be able to invoke additional capabilities without waiting for specific compilers and libraries to be updated by multiple providers. Once oneAPI is more widely adopted, it then becomes possible to significantly lower the cost of application development and deployment, noted Reinders.
It’s not clear whether the oneAPI initiative will become more widely adopted, either under the auspices of Intel or by some type of industry consortium that provides a neutral venue for building a common set of compilers and libraries. The one thing that is certain is that DevOps teams are vested in trying to standardize compilers and libraries across as many platforms as possible. The challenge, of course, is making sure developers only invoke compilers and libraries that can span multiple classes of processors versus ones that may have been optimized for a specific type of processor architecture.
