Igor’s Lab — “KNOD shifts network processing to AMD GPUs”

Hardware-press report (igorslab.de, English edition) on knod, an RFC patch series that offloads Linux network processing to AMD GPUs. Secondary coverage of a kernel mailing-list posting, so T3: no code read, no benchmarks run, and the original thread is not linked in the excerpt fetched.

What it reports

• KNOD = “in-kernel network offload device.” The NIC DMAs packets into GPU memory; the GPU processes many in parallel.
• “The Linux kernel itself manages the GPU queues, compiles the programs via JIT into suitable GPU machine code, and then initiates their processing.”
• “A userspace runtime such as AMD ROCm or CUDA is not intended to be a permanent part of the data path.” AMD is the practical target because its Linux driver is mainline open source.
• Offload candidates: XDP programs, IPsec Security Associations, load balancing and filtering.
• Tested on GCN and RDNA2 GPUs.
• Status: RFC on the kernel mailing list as of 2026-07-22, “not yet a fixed Linux version for the inclusion of KNOD.”

What it does not report

No developer or employer is named. No kernel version. No latency, throughput, or CPU-utilization numbers, and no reaction from kernel maintainers — the two things that would show whether the idea survives review. Treat the design as reported and the value as unmeasured; a primary read of the LKML thread would upgrade this.

Why it is here

It extends the spoke’s kernel-programmability line (ebpf → sched-ext → knod) to a new execution substrate, and it touches the data path operators run in production. See knod for the mechanism and the DPU adjacency.