Presented at Kubernetes Community Days (KCD) Czech & Slovak 2026.
Abstract
GPUs are the new black box. As AI workloads take over production, SREs are left blind: the observability stack stops at the CPU boundary. When an inference job degrades during an on-call incident, there is little to look at on the GPU side.
This talk explores how eBPF extends observability past that boundary:
- GPU runtime tracing with zero code changes
- Injecting eBPF into GPU kernels
- Current research directions in AI infrastructure monitoring
- Three observability layers: GPU runtime, inference-aware, and infrastructure correlation
- Tools to try today: GPUprobe, eInfer, eACGM, and the emerging eGPU work