Back

CudaCKPT

2025Featured

CudaCKPT is a checkpoint/restore runtime for single-GPU CUDA processes. It snapshots device memory, execution context, streams, and kernels so long-running GPU jobs can resume without recomputing from scratch, with measured <4% runtime overhead and 1.8GB/s checkpoint throughput.

Highlights

  • -Transparent capture of GPU memory, context, streams, and kernels for single-GPU CUDA processes
  • -95% reduction in recomputation on restored jobs, with under 4% runtime overhead
  • -1.8GB/s checkpoint throughput on the restore path
C++CUDAGPUSystems
← All projects