Tech • AI • Robotics • Game
NVIDIA DOCA Skills vs No Skills: Building RDMA on NVIDIA BlueField-3 with Fewer Commands & Less Code
7/10TL;DR
Testing on a BlueField-3 showed that NVIDIA DOCA skills let an AI agent build and verify a real RDMA application with far fewer hardware commands and much less handwritten code.
KEY POINTS
Same target, different path
Two AI coding agents were given the same task: build a Go program that uses NVIDIA DOCA to send real RDMA traffic on a BlueField-3 and verify that the payload arrived. Both agents ultimately produced a working DOCA RDMA application and confirmed the bytes on hardware. The difference was not the end result, but how much work was required to get there.
Skills reduced hardware probing
The agent without DOCA skills had to discover the correct module, headers, build settings, and sample code through repeated trial and error on the device. That process led to roughly 37 commands run on the hardware. The skilled agent completed the same task in about 20 commands, or close to half as many.
Direct use of NVIDIA-supported code
With the DOCA RDMA and DOCA Programming Guide skills available, the agent was directed immediately to NVIDIA’s shipped DOCA RDMA sample instead of searching manually. The skills also supplied the correct CGO link flags, pointed to the installed DOCA version, and established the rule that the solution should not fall back to raw verbs. That shifted the effort from rediscovery to adaptation.
Far less handwritten code
Without skills, the agent wrote about 695 lines of low-level DOCA code by hand. With skills, it reused NVIDIA’s tested sample and added a 189-line wrapper. That represents roughly 73% less handwritten code while still reaching the same verified hardware result.
Verification happened on real hardware
The test was framed around an actual BlueField-3 environment rather than a mock or simulated setup. In both runs, the resulting application sent real RDMA traffic and verified that the payload was received. The measurable gain came from fewer steps and less custom glue code, not from skipping validation.
Benefits extended beyond RDMA
The same pattern was reported in other DOCA workloads, including Flow, COM channel, Ethernet, DMA, and SHA. In cases where the initial prompt did not explicitly mention DOCA, the presence of skills had a larger effect: without them, one run reportedly used a Linux kernel packet path and linked no DOCA at all, while the skilled run built a real DOCA Flow pipeline and read hardware counters from the card.
Supported samples replaced reinvention
For broader networking tasks, the contrast was again framed as reuse versus reinvention. One path relied on hundreds of lines of raw low-level code, while the skilled path wrapped NVIDIA sample implementations with a much smaller adapter layer. In one comparison, a raw implementation of about 660 lines was replaced by an 87-line wrapper around shipped sample code.
Measured effect was consistency, not a one-off
The reported takeaway was not a single isolated win but a repeated pattern across runs. Where an agent already chose DOCA, such as in DMA and SHA, the skills did not appear to hurt outcomes. Across tests, they consistently reduced command count, cut custom code, and kept the solution aligned with NVIDIA-supported implementations.
CONCLUSION
On BlueField-3, the main impact of DOCA skills was operational efficiency: fewer hardware interactions, less handwritten low-level code, and stronger reuse of NVIDIA reference implementations. The reported results suggest that, when building DOCA applications in Go, targeted platform knowledge can materially shorten the path to a verified working result.
Ask a question