KNOD: Offloading Linux network packet processing to AMD GPUs

KNOD (In-Kernel Network Offload Device) is an experimental Linux kernel framework that repurposes consumer-grade AMD GPUs to handle high-rate packet processing, addressing the CPU bottleneck at 100-400 Gbps link speeds. It maps XDP-style packet programs into GPU-executable code via kernel-level JIT compilation, lets NICs DMA packets directly into GPU-accessible memory, and runs identical packet operations across thousands of GPU lanes in parallel. By keeping the data path inside the kernel and avoiding userspace runtimes like CUDA or ROCm, KNOD reduces latency and CPU overhead; early benchmarks show up to 70 million packets per second on a Katran-derived workload using an AMD Radeon PRO V620.

The current implementation targets AMD because of the open-source amdgpu driver integration in mainline Linux and hardware features (RDNA2-era support for 64-bit atomics) that enable the needed memory and execution primitives. That focus makes KNOD a practical alternative to expensive SmartNICs or DPUs for dedicated-server workloads by freeing CPU cores otherwise pinned for polling, improving server ROI, and reducing vulnerability to traffic spikes. KNOD remains experimental: broader hardware support, APU tradeoffs, motherboard compatibility, and production-readiness need resolution before widespread deployment, but the approach demonstrates a viable path to GPU-accelerated packet filtering and L4 operations inside the kernel.