https://github.com/NVIDIA/cuda-samples/tree/master/Samples/UnifiedMemoryPerf
Unified and other CUDA Memories Performance
该示例演示了使用带/不带提示的统一内存矩阵乘法内核以及其他类型的内存(例如零拷贝缓冲区,可分页,分页锁定的内存,在单个GPU上执行同步和异步传输)的性能比较:
UMhint UMhntAs UMeasy 0Copy MemCopy CpAsync CpHpglk CpPglAs
"UMhint", // Managed Memory With Hints
"UMhntAs", // Managed Memory With_Hints Async
"UMeasy", // Managed_Memory with No Hints
"0Copy", // Zero Copy
"MemCopy", // USE HOST PAGEABLE AND DEVICE_MEMORY
"CpAsync", // USE HOST PAGEABLE AND DEVICE_MEMORY ASYNC
"CpHpglk", // USE HOST PAGELOCKED AND DEVICE MEMORY
"CpPglAs" // USE HOST PAGELOCKED AND DEVICE MEMORY ASYNC
测试结果:
-
(Dell Precision 5520) Device 0: "Quadro M1200" (Maxwell cc5.0)
Quadro M1200
-Jetson Xavier capability 7.2 (Volta)
- (机械革命S1):mx150(Pascal)
网友评论