KNOD Progressing For Linux In-Kernel GPU Offloading For Network Packet Processing

A while back we covered KNOD as an experimental solution for in-kernel network offloading to AMD GPUs. The open-source KNOD project allows leveraging GPUs to handle packet processing in a very performant and energy efficient manner without needing any specialized hardware. KNOD has continued and a development update was shared this week at the Linux Plumbers Conference in Prague. KNOD continues showing a lot of interesting potential for exploiting GPUs as a kernel-managed packet-processing accelerator for networking. KNOD will JIT compile an XDP program into GPU machine code, the network adapter will place packets directly into GPU-accessible memory, and the GPU will process them. RX IPsec abilities were recently added beyond the original XDP programs. KNOD has proven performant with getting up to 70 million packets per second for XDP workloads and 80 Gbit/s for RX IPsec on consumer GPU hardware. The developers behind KNOD find it to be very cost effective as well as being more power efficient than just using the CPU for packet processing. While AMD GPUs have been their main target to date, it was mentioned at LPC 2026 they are going to explore Intel Arc graphics support too as well as seeing that KNOD will work with the likes of Broadcom and Intel NICs too. AMD integrated graphics will also be explored rather than just discrete GPUs. Those wishing to learn more about the interesting KNOD concept can do so via the LPC presentation assets.