NVIDIA DOCA Skills vs No Skills: Building RDMA on NVIDIA BlueField-3 with Fewer Commands & Less Code
NVIDIA DOCA Skills reduce AI agent handwritten code by 73% and hardware commands by 50%
“73% less handwritten code and it reached the same verified result in about half the commands on the hardware.”
NVIDIA demonstrated that pre-installed DOCA Skills for AI coding agents dramatically reduce the effort to build RDMA applications on BlueField-3 DPUs — cutting handwritten code from 695 to 189 lines and hardware probes from ~37 to ~20 commands. The skills work by giving agents authoritative module names, build flags, and NVIDIA-shipped samples upfront, eliminating trial-and-error discovery. This is a concrete example of domain-specific agent context (skills/knowledge packs) delivering measurable efficiency gains in infrastructure programming tasks.