NVIDIA NVIDIA
September 30, 2026

NVIDIA DOCA Skills vs No Skills: Building RDMA on NVIDIA BlueField-3 with Fewer Commands & Less Code

This demo is a side-by-side comparison of two AI coding agents tasked with building a Go-based RDMA application using NVIDIA DOCA on a BlueField-3 platform.

Both agents are given the same goal: implement a working RDMA data path and verify that bytes successfully arrive on hardware. The key difference is that one agent has access to NVIDIA DOCA “skills,” while the other must discover the required modules, headers, and build process from scratch through iterative hardware probing.

The experiment highlights how access to curated domain knowledge impacts efficiency
The “skilled” agent rapidly identifies the correct DOCA modules, samples, and CGO build configuration.
The “unskilled” agent relies on trial-and-error, repeatedly probing hardware and discovering components incrementally.
Both approaches ultimately produce a working, hardware-verified RDMA application.

Key results from the comparison
~73% reduction in handwritten code when using DOCA skills
~50% fewer hardware interaction commands
Faster convergence to NVIDIA-supported sample-based implementation
More consistent use of validated DOCA pipelines instead of low-level reimplementation

The demo illustrates how structured platform knowledge can significantly reduce development effort while maintaining correct and verifiable low-level hardware behavior on accelerated networking systems.
72

Comments 0

Replying to
5000 max

Be the first to comment!

This website uses cookies
This website uses cookies.
To maintain our services, we use cookies. They allow us to personalize ads, analyze our traffic, and provide social media features.

Customize your preferences

Select the cookies you accept: