Posted: April 2, 2021
I'm pleased to announce that the Intel® oneAPI Toolkits are updated for the first time since their inaugural production release in late 2020, offering additional performance, new capabilities for compute, AI, media, ray-tracing, HPC, and more. It is worth updating your oneAPI installations. These toolkits represent the most complete and diverse set of oneAPI-initiative compliant tools (compilers, libraries, pre-optimized frameworks, analyzers, debuggers) available today. The Intel oneAPI toolkits are prebuilt, ready to install, and supported by Intel engineers via community forums and Priority Support.
Intel® oneAPI Toolkits help us use a unified approach to developing applications and solutions that harness CPUs, GPUs and FPGAs, together. Some toolkits focus on native code development using Data Parallel C++ and Fortran. AI and Data Science toolkits enable developers using machine learning and deep learning techniques through Python and AI frameworks. All the toolkits help us unleash the power of heterogeneous computing systems by opening access to the processing power in our systems.
What's in the oneAPI toolkit updates
This release (2021.2) ranks as a 'dot' update release but still manages to adds enough useful enhancements to make it worthwhile to install the update.
The projects that make up the toolkits are continually active, and this update adds a generous list of new and enhanced features. You can get the updates in the same places as you got your original downloads - the oneAPI toolkit site, via containers, apt-get, or use directly in the Intel® DevCloud. Instructions for all these are on the Intel oneAPI Toolkits page.
If you care about top performance, it is always worth keeping up-to-date. Long-time users will know that Intel never misses an opportunity to tune compiler optimizations and enhance its libraries, and this is no exception. In reviewing the release notes, it is unsurprising to see that many optimizations are for recent Intel® processors, and Intel® Xe graphics (GPUs). Whether you depend on accelerated frameworks, compilers, libraries, or other tools, the update is likely to be useful.
AI and HPC enhancements
AI customers will see numerous enhancements for XGBoost and scikit-learn performance, plus upgrades to Intel® Optimization for Tensorflow and PyTorch. PyTorch provides improved INT8 optimizations, while the Intel® Low Precision Optimization Tool extends support to new backends such as PyTorch and ONNX Runtime. These both strongly support effective quantization that many find valuable for increasing performance. Intel® oneAPI Deep Neural Network Library (oneDNN) introduced binary post-op APIs for (de)-convolution, pooling, eltwise, binary, inner product, matmul and reduction (GPU only) along with performance optimizations for CPUs and GPUs.
The Intel® MPI Library expanded container support, and updated Mellanox OFED support. Intel® Advisor enhanced the Source View analysis for Offload Modeling and GPU Roofline capabilities. To let us examine hardware utilization at a glance, Intel® VTune™ Profiler added a Platform Diagram, a new starting point for the Input and Output analysis. It reveals system topology and high-level utilization metrics for hardware resources including PCIe devices, Intel® Ultra Path Interconnect, and memory.
Media acceleration, Ray-Tracing, and Rendering continue to delight
Volume ray tracing enhancements include VDB multi-attribute volume support, faster structured and VDB volume sample and quicker interval iteration on structure, VDB, and unstructured volumes. An enhanced denoising library now supports directional lightmaps and Apple Silicon. Colored transparency, colored shadows, cones, and cylinder geometries are available within the ray tracing API and engine help with scientific visualization and general rendering.
Library GUI enhancements include an improved Lights editor for easy scene lighting, gITF loader improvements enabling animation and skinning of triangle meshes, lights, materials and textures, and better scene file load+save by including model transforms, material, lights and camera state.
Media support for fast transcoding, streaming, and more, see multiple enhancements including the Intel® oneAPI Video Processing Library (oneVPL) expanding support for Intel GPUs including Intel® Iris® Xe and Iris® Xe MAX Graphics when using Core Media APIs (these were previously in the Intel® Media SDK and are now bundled in oneVPL for a more seamless oneAPI experience). Backwards compatibility provides developers a migration path from Media SDK to new features and capabilities, such as easier device enumeration, easier video processing initialization and extended AV1 encode features.
XPU programming with attitude
To dive into the details that motivates oneAPI, SYCL, and DPC++, I'm launching my inaugural blog on heterogeneous programming (#xpublog). I start with an introduction to XPUs, DevCloud, and SYCL/DPC++. It's a modest start, and we'll grow from there. Please check it out and leave feedback on what you would like to hear more about.
Another great place to learn is IWOCL / SYCLcon, and the pre-conference oneAPI summit "at IWOCL21". The SYCLcon talks are available to view on demand starting April 26; free registration will get you all the information on how to view, simply visit the IWOCL21/SYCLcon21 registration page. I'm honored to be giving a talk titled "SYCL, DPC++, XPUs, oneAPI – a View from Intel."
Learn how to make the most out of the Intel oneAPI toolkits
There continues to be a huge amount of training available on all aspects of the toolkits, and DevCloud remains a fantastic place to try out the tools without having to set up your own system. It is a great place to test code and workloads across all types of Intel CPUs and accelerators for free. DevCloud is loaded with the new tools as well - so they are there for you to enjoy as well. If you don't know about DevCloud yet, I recommend you check out my new blog series (#xpublog).
Update now for all the best in oneAPI
I barely scratched the surface, but the message is clear: update to get the best that Intel® oneAPI Toolkits have to offer and check out the release notes for your favorite components to see what new support may prove useful for you.
Visit my blog for more encouragement on joining in the movement to heterogeneous programming that is possible with oneAPI.
It's an attitude, and a journey - to help all software developers.