Accelerators of various kinds are increasingly important to enable general purpose processors to support very demanding applications such as Cloud-RAN baseband processing or large-scale AI/ML. Intel has set out a vision of a single programming model and API framework which would work across all kinds of accelerator chip technologies, such as FPGAs or graphical processing units (GPUs).
It calls its vision XPU and it has announced the gold release of oneAPI, the foundation for its strategy.
“All of those architectures are not supported with the same languages, libraries, and tools,” said Jeffrey McVeigh, general manager of visual computing products at Intel, who also announced the company’s first GPU for data center and other large-scale environments. “This makes it very difficult to be able to reuse code and be productive across those architectures.”
Currently, software developed to run on a CPU has to be reworked to take advantage of GPU or FPGA acceleration. OneAPI aims to address that by acting as a low level abstraction layer that enables developers to write once and run on anything. There is also a common language called Data Parallel C++, based on an evolution of OpenCL and combined with optimized libraries for specific workloads, such as deep learning.