eBPF GPU Tracing: Finding What nvidia-smi Misses

GPU metrics lie. nvidia-smi says 97% utilization while your model stalls. This series traces what actually happens inside GPU workloads — CUDA kernel launches, graph captures, memory allocations, scheduling decisions — using eBPF uprobes on libcudart.so. Each post is a real investigation with reproduction steps and an open-source investigation database you can query yourself.