I profiled my application using the command
nvprof --metrics all ./app &> prof.txt
. The thing I don’t understand is that why the number of device memory read transactions is greater than that of L2 read transaction.
1 dram_read_transactions Device Memory Read Transactions 488766156 488766156 488766156
1 l2_read_transactions L2 Read Transactions 268076721 268076721 268076721
My understanding is that device memory read transactions only happen when there are L2 misses. From this standing point of view, the number of L2 read transactions should be greater than the number of device memory read transactions. Could someone explain to me why profiling contradicts to my understanding?