Data was measured using
http://git.ipe.kit.edu/matthias/gpuperf
and the following commands:
$ ./perf --source-size=4096:25165824:4096 --dest-size=134217728 > result.128 $ ./perf --source-size=4096:25165824:4096 --dest-size=67108864 > result.64 $ ./perf --source-size=4096:25165824:4096 --dest-size=33554432 > result.32
The columns are
block size | destination size | wall time | event time | throughput
The throughput is estimated using the wall time.