NVIDIA's Deep Learning Accelerators, NVDLAs, Could Receive Mainline Support in the Linux Kernel

Using NVIDIA's Deep Learning Accelerator (NVDLA) technology could become easier, as a proposal to merge driver code into the Linux kernel for mainline support is submitted to the development list.

The NVDLA, designed for high-efficiency acceleration of machine learning and artificial intelligence workloads with a view to use in embedded and Internet of Things (IoT) applications, is open source — released by the company back in September 2017 under the custom NVIDIA Open NVDLA License. Now in its second-generation release, as found on-board the Jetson AGX Orin module, it's a powerful system — but one which requires software support to be of use.

Like the company's graphics products, using NVDLA hardware has typically required the installation of drivers provided by NVIDIA — but a new submission to the Linux kernel, brought to our attention by Phoronix, could mean mainline support is on the way.

"The NVIDIA Deep Learning Accelerator (NVDLA) is an open source IP which is integrated into NVIDIA Jetson AGX Xavier," developer Cai Huoqing writes in the patch submission that adds the required code, "so add driver support for this accelerator."

It's not immediately clear whether the driver submitted in the patch set will support both the original NVDLA as launched with the Xavier architecture and the new NVDLA v2 found in NVIDIA's Ampere architecture; it's worth noting that the company's GitHub repository for the project hasn't been updated in four years and still only includes NVDLA v1.

Kernel developer Christian König, employed by NVIDIA rival AMD, commented on the patch to say it "doesn't looks so bad on first glance (regarding coding style etc..)," but there has not yet been an official notification that the patches have been or will be accepted into the mainline kernel.

More information is available on the Linux Kernel mailing list archive.