Does moving a GPU server's request loop into the kernel save CPU? | eunomia eunomia.dev

Eunomia’s bpf-gpu-server prototype moves a GPU server’s TCP request loop into eBPF and io_uring, but tests on an RTX 5090 found CPU use and latency broadly matched an efficient epoll server using the same GPU executor. The kernel path cut user time but shifted work into system time; the post says most per-request cost remained in Linux’s TCP stack, while capacity and TCP_NODELAY had much larger effects on latency.