cedana / benchmarks · measured GPU performance
Performance measured
in production conditions.
Reproducible GPU cold-start results, full methodology, and the configurations behind every number.
All benchmarks · 02
measured() on real GPU systems.
Native startup compared with Cedana restore, with the methodology and configuration published alongside the result.
Frontier LLM cold start: native vs. Cedana restore
Measured across MiniMax, GLM, Kimi, and DeepSeek models. Cedana restores every model in 57–70 seconds, with a peak 29.3× speedup.
view benchmark →FLUX diffusion cold start: 50 minutes to 73 seconds
Per-resolution torch.compile across 11 fixed resolutions pushes native startup to roughly 50 minutes. Cedana restores the compiled pipeline in 73 seconds.
view benchmark →