Open Source Summit Korea 2026

GPU compute resources remain inaccessible to the Linux kernel without userspace runtimes (ROCm, CUDA), forcing network packet processing to stay CPU-bound regardless of available GPU capacity.
We present KNOD (in-Kernel Network Offload Device), a Linux kernel framework that enables GPU-accelerated network packet processing entirely within the kernel. KNOD compiles and launches GPU kernels natively, dispatching packet processing workloads directly to the GPU — no ROCm, no userspace involvement, using only publicly documented hardware interfaces.
XDP offload is KNOD's first use case. Building on our LPC 2025 talk, this talk covers: in-kernel GPU kernel compilation targeting AMD GFX9/Vega ISA with post-processing optimizations (latency: ~118µs → ~46µs); a lock-free GPU-side queue between NIC RX path and GPU compute; and branch divergence mitigation to maximize SIMT utilization across heterogeneous packet flows.
At equivalent throughput (~32 Mpps), KNOD reduces system-wide CPU utilization from 40% (CPU-based XDP) to 17-20%. We discuss lessons from MACsec/WireGuard offload attempts and the path toward KNOD as a general kernel framework for GPU-accelerated network functions.