Linux Driver Prepared For The Qualcomm Cloud AI 100 Accelerator

Announced over one year ago was the Qualcomm Cloud AI 100 Accelerator as designed for offering superior AI inference performance in the cloud. Since then not much has been heard of this accelerator but their engineers did publish a Linux driver on Thursday.

At least as of April 2019, Qualcomm believed their Cloud AI 100 offering could deliver 10x the perf-per-Watt of other inference solutions and would support PyTorch, TensorFlow, Keras, ONNX, and all the other popular deep learning solutions.

Qualcomm Cloud AI 100 was expected for H2'2019 while at least as far as upstream Linux kernel support is concerned the first patches were only published this week. In fact, while checking for any other recent activities on the Cloud AI 100 front I may have missed, most web references to it are dated back just from April 2019 when first announced.

A set of eight Linux kernel patches were posted on Thursday for enabling this PCI Express adapter card under Linux. Qualcomm's Jeffrey Hugo says their goal is to have "the product fully upstream" in the kernel so they began with an initial "request for comments" on the driver.

The Qualcomm Cloud AI driver is dependent on the recently introduced MHI bus for communication with the deep learning specialized SoC. The communication between kernel and user-space is done using a char device interface.

The initial kernel patches for supporting the Qualcomm Cloud AI accelerator can be found on the kernel mailing list. The patches are still in flux with additional feature work pending. At this point no open-source user-space patches appear to have been posted.