Modern applications in areas like Machine Learning, Artificial Intelligence, and 3D Graphics, require a synergistic software/hardware ecosystem that allow developers to take full advantage of hardware accelerators. In this scenario, it is critical to have a low-level API that can easily support and adapt to any device, in order to minimize the impact in upper-levels of the software stack when exposing novel hardware capabilities to higher-level programming models and frameworks.
Level-Zero API, part of Intel OneAPI product, defines a device-independent, vendor-agnostic, low-level, direct-to-metal interface to accelerator devices that abstracts users and upper-level components of the software stack from the specifics of the target devices, while providing them with the access needed to fully exploit their hardware capabilities. This is essential for Intel to expose new hardware features at a faster pace and to effectively compete against established CUDA-based ecosystem from NVIDIA.
This presentation offers an overview of the rich set of interfaces defined in Level-Zero, focusing on capabilities such as unified-shared memory, peer-to-peer communication, and inter-process communication. Additionally, the status of the implementation of Level-Zero and its adoption by higher-level compiler, analysis tools, performance libraries and other frameworks are presented.