本演示文稿深入探讨了使用CUDA技术在GPU上管理内存的最佳实践与策略,旨在帮助计算机科学家和工程师优化高性能计算应用。
Bob Crovella, 6182020CUDA UNIFIED MEMORY
AGENDA
- Managed Memory - basic idea, objectives, benefits
- Demand-Paging, Oversubscription, Concurrency, Atomics
- Use Cases: Deep Copies, Linked Lists, C++ Objects, Graph Traversal
- Performance: Prefetching, Hints
- Multi-GPU Considerations
- Further Study
- Homework
THE CUDA 3-STEP PROCESSING SEQUENCE
Recall from Module 1…
Wouldn’t it be nice if we didn’t have to do (i.e. write the code for) steps 1 and 3?
INTRODUCING UNIFIED MEMORY WITH DEMAND PAGING