Linux Patches Introduce "KNOD" For In-Kernel Network Offloading Directly To AMD GPUs - Phoronix

What the report says

Phoronix reports that a new set of Linux kernel patches introduces “KNOD,” short for in-kernel network offload device, aimed at moving network packet processing onto GPUs. The patches were posted to the Linux kernel mailing list on Sunday by Taehee Yoo as a Request for Comments, following an earlier concept presentation at Linux Plumbers Conference in Japan. The initial work focuses on AMD GPUs.

According to Phoronix, KNOD is designed to operate entirely inside the kernel, without depending on user-space GPU frameworks such as AMD ROCm. The proposed kernel code would manage GPU queues, compile per-packet programs into GPU machine code, and dispatch processing work directly. The approach would allow network adapters to place received packets straight into GPU memory via DMA, with the GPU handling batches of packet work in parallel.

The patches are intended to support offloading existing Linux networking mechanisms, including XDP programs and IPsec security associations, in a way that is transparent to users of those paths. The broader goal is to reduce CPU pressure in high-throughput networking tasks such as load balancing or cryptographic packet handling, where per-packet work can otherwise consume many host CPU cores.

Phoronix says development has been carried out on AMD GCN hardware, including Radeon RX Vega, as well as RDNA2 systems. The article notes AMD’s open and upstream driver stack as an advantage for this early work, while NVIDIA and Intel GPU support are not part of the current implementation. As an RFC submission, KNOD is at an early review stage and would still need discussion and acceptance before any possible inclusion in the Linux kernel.