NVLink-C2C Removes the Copy, Not the Synchronization

The CUDA Programming Guide assumes the host and the device keep separate memories with a copy between them. My DGX Spark doesn’t. So I went and learned what NVLink-C2C actually is compared to PCIe, how coherence behaves across it, and what that changes for someone writing C++ on this machine.

CUDA Performance, Part 1: Device Busy Is Not GPU Busy

A program can report 98.8% device utilization while leaving its compute units idle 39% of the time. This is what I learned building a TUI for nsys timelines, and why the second number is the only one that matters.