Qualcomm Publishes Open-Source Compiler & User-Space For Their Cloud AI Accelerator
Last month Qualcomm published updated patches for their Cloud AI 100 kernel driver to support this inference accelerator. The Qualcomm engineers said at the time that their user-space driver and associated compiler would be published shortly. That panned out and the user-space portion of this open-source AI inference stack was recently published.
Qualcomm has out their QAIC user-mode driver and compiler for interfacing with their work-in-progress kernel driver that is working its way to the kernel's new "accel" accelerator subsystem.
Via this GitHub repository is the user-mode driver / software kit for the Cloud AI 100 and this repository has the compiler. This is crucial for getting the kernel accelerator driver mainlined as open user-space clients to interface with the kernel driver are mandatory. This milestone is similar to Intel's Habana Labs who for several years has also provided a fully open-source kernel and user-space stack for their AI hardware.
Both Qualcomm's user-mode driver and the compiler are published under a BSD-3-Clause-Clear license. The software kit also includes a few sample applications, including a tool for querying QAIC 100 cards and for deploying applications to the hardware.
The Qualcomm Cloud AI 100 features up to 16 AI cores on a 7nm process, up to 32GB of LPDDR4x system memory in a quad channel configuration., and is designed for AI inference from the cloud to the edge.
Now that the open-source user-space code is available, we'll see if that helps get the "qaic" accelerator kernel driver mainlined soon.