Snowflake's Semi-Persistence Cuts Multi-Model GPU Sleep/Wake Cycles by Up to 19.9x

StasBekman · x · 2026-09-01

Snowflake AI Research's Semi-Persistence tackles multi-model serving on shared GPUs: keeping every model loaded wastes capacity, but swapping only helps if it stays out of the serving path.

The approach keeps model weights in a pinned CPU-memory pool and treats the GPU copy as temporary; when demand returns, weights stream back over PCIe and NVLink in parallel. Across models from 2B to 397B parameters, internal benchmarks showed 5.6x–19.9x faster sleep/wake cycles vs. the baseline vLLM path, with sub-second swaps for single-GPU models.

Related event: Snowflake's Semi-Persistence Speeds Up Multi-Model GPU Swapping Up to 19.9x(2 posts)→

Original post →

More from Infra

Infra channel →