Open-Source Driver for VeriSilicon NPU Merged into Mesa
Open-source driver support for VeriSilicon’s Vivante Neural Processing Unit (NPU) has been merged into the Mesa graphics library. This development continues the trend toward open, community-supported AI acceleration hardware. The availability of open-source drivers can be a key enabler for using such hardware in custom avionics where closed-source firmware presents certification challenges.
- The open-source driver for VeriSilicon's NPU is integrated into Mesa through a new TensorFlow Lite delegate called Teflon. This allows the driver to be automatically packaged and distributed with major Linux distributions like Debian, Ubuntu, and Fedora. The work was sponsored by Libre Computer and spearheaded by developer Tomeu Vizoso. - This driver is part of the broader Etnaviv project, which focuses on reverse-engineering and developing open-source drivers for Vivante GPUs and NPUs. The NPU architecture was found to be similar to existing Vivante graphics cores, which simplified the development process. - Performance of the open-source driver is a key focus, with ongoing optimizations to better utilize the hardware. Early performance for some workloads was about three times faster than CPU-based inference and is continually improving, though still trailing the proprietary driver in some cases. Recent updates have included optimizations like weight compression and improved caching, leading to significant performance jumps. - VeriSilicon's Vivante NPU IP is designed for a range of applications from low-power IoT and wearable devices to automotive and data processing. The company has shipped over 100 million AI chips incorporating their NPU IP. - The Mesa 3D graphics library is an open-source implementation of various graphics APIs, including OpenGL, OpenGL ES, and Vulkan. It provides a standardized layer between applications and the underlying hardware drivers for a wide variety of GPUs and accelerators. - Open-source drivers are critical for the adoption of hardware in systems requiring high levels of scrutiny and customization, such as in the aerospace industry. The transparent nature of open-source software can simplify the certification process for standards like DO-178C, which governs software in airborne systems. - Google has collaborated with VeriSilicon to open-source the Coral NPU IP, which is based on the RISC-V instruction set. This initiative aims to foster a more robust ecosystem for edge AI development on an open platform. - The Gallium3D framework within Mesa is instrumental in this effort, as it separates hardware-specific drivers from the front-end APIs. This modularity makes it easier to add support for new hardware like the VeriSilicon NPU and potentially other accelerators in the future.
