inference/research
verifiedhigh · 0.90updated 2026-10-03

Getting Started with CUDA Graphs

blog

Fields

id
sources/flop-nvidia-cuda-graphs
Publisher
NVIDIA Technical Blog (Alan Gray)
Kind
blog
Published
2019-09-05
Accessed
2026-10-03
notes
Documents CUDA Graphs: capture a launch sequence once and replay it, removing per-kernel launch overhead. This is the mechanism engines reach for at batch 1, where per-op launch latency is comparable to the work itself. Also the reason batching engines want static shapes.

Notes

Documents CUDA Graphs: capture a launch sequence once and replay it, removing per-kernel launch overhead. This is the mechanism engines reach for at batch 1, where per-op launch latency is comparable to the work itself. Also the reason batching engines want static shapes.

Cross-references

This record references no other records.