OneAPI is Intel's application programming interface for XPU hybrid computing, which can simplify the difficulty of cross-architecture program development, and the new oneAPI 2022 toolkit can also bring better execution efficiency.
Intel said that according to the Evans Data survey, 40% of heterogeneous system developers use more than 1 processor, processor core or coprocessor. Through oneAPI, it can provide a single and complete solution across processors (CPUs), graphics processors (GPUs), FPGAs and other architectural computing units and accelerators, greatly reducing the development threshold of heterogeneous computing, allowing developers to freely choose the best hardware for specific solutions without being limited by the economic and technical burdens brought about by the dedicated program development model.
Intel's latest oneAPI 2022 toolkit integrates the entire program development tool, including compilers, libraries, pre-optimized frameworks, analyzers and debuggers, and has added and enhanced more than 900 features in the past 1 year, including the world's first CPU and GPU C, SYCL and Fortran through the universal LLVM backend, data parallel Python unified compiler, advanced accelerator performance model and tuning, As well as AI and lighting focus on visualizing workload performance acceleration.
In addition, the DPC Compatibility Tool automatically transfers 90% to 95% of CUDA program code to SYCL/DPC, allowing developers to increase the productivity and rate of code development when creating cross-architecture applications.
OneAPI can reduce the difficulty of developing hybrid architecture computing programs.
OneAPI supports a variety of processors (CPUs), drawing processors (GPUs), FPGAs and other architectural computing units, accelerators.
oneAPI integrates components such as compilers, libraries, program frameworks, middleware, optimized applications, and more.
Optimized for the latest and upcoming hardware, oneAPI delivers a more robust overall performance output through avX-VNNI on the 12th generation Core processor codenamed Alder Lake, Advanced Matrix Extension (Intel AMX) for the next-generation Xeon Scalable processor codenamed Sapphire Rapids, and the upcoming Xe GPU.
In terms of AI performance optimization, deep learning framework performance is up to 10 times faster than previous versions through the latest Intel Optimization for TensorFlow and Intel Optimization for PyTorch. The new Intel Extension for Scikit-learn performs on Intel CPUs, which can speed up machine learning algorithms by more than 100 times compared to the open source version. Intel Neural Compressor delivers better AI inference performance through post-training optimization techniques across multiple deep learning frameworks.
The new version is also deeply integrated with more development environments, allowing users to develop programs in familiar environments such as Microsoft Visual Studio Code, Microsoft Visual Studio 2022, WSL2, and more.
In addition, advanced light attention features include cone telemetry, auxiliary features for noise reduction, and support for the FP16 data type, providing a more solid form factor and reduced rendering time, improving overall rendering performance. Support for upcoming Xe GPUs and instant noise cancellation for more realistic image quality.
For example, the Embree path attention technology can simulate the transmission of light in a 3D environment, and the xPU path tracer created by the oneAPI Rendering Toolkit library, and the AI denoiser Open Image Denoise bring sharper image quality.
The video above shows the technology of Embree and Open Image Denoise, and you can see how the Xe architecture performs with oneAPI for the system.
Readers can access the latest oneAPI toolkit in the Intel Developer Zone, as well as files through containers, apt-get, Anaconda, etc., or accessed by the Intel DevCloud cloud development environment.