GVisor GPU cold starts cut using memory snapshots for CUDA workloads

Engineers at Cerebrium have described a technique for reducing GPU cold start times in GVisor by taking memory snapshots of CUDA workloads. The approach allows a ready compute environment to be restored in seconds rather than minutes. The solution is particularly useful for machine learning applications that require rapid container scaling.