Unified Memory in CUDA¶

1. What primary benefit does CUDA unified memory offer for memory management between CPU and GPU?

• A. It enables CPU and GPU to share memory without explicit data transfers
• B. It increases GPU memory size
• C. It automatically optimizes code performance
• D. It allows direct access to GPU registers from the CPU

Using cudaMallocManaged() vs __managed__ Variables¶

2. Which of the following functions allocates unified memory in CUDA?

• A. cudaMalloc()
• B. cudaMemcpy()
• C. cudaMallocManaged()
• D. cudaFree()

Example: Without Unified Memory¶

3. In a non-unified memory setup, which operation is required to transfer data from CPU to GPU?

• A. cudaMalloc()
• B. cudaMemcpy()
• C. cudaDeviceSynchronize()
• D. cudaFree()

Example: With Unified Memory¶

4. True or False: When using unified memory, we no longer need to call cudaMemcpy() to transfer data between the host and device.

Prefetching Pageable Memory¶

5. What is the purpose of cudaMemPrefetchAsync() in CUDA unified memory?

• A. To allocate unified memory
• B. To free unified memory
• C. To move memory to a specific device in advance, reducing page faults
• D. To synchronize memory between CPU and GPU

Benefits of Using Unified Memory¶

6. Which of the following is NOT a benefit of CUDA unified memory?

• A. Simplified code without explicit memory copy operations
• B. Automatic memory paging between CPU and GPU
• C. Guaranteed best performance in all applications
• D. Easier management of memory resources between CPU and GPU