Looking at the code, it really isn’t very suprising that the GPU is slower. The memory access patterns are about as suboptimal as you could possibly make them for a compute 1.1 device. There is a section in the programming guide which discusses the concept of memory coalescing and how to achieve peak memory throughput on the GPU. You might want to review it.
Related topics
| Topic | Replies | Views | Activity | |
|---|---|---|---|---|
| GPU/CPU precision comparison and Kernel instructions question | 5 | 679 | April 4, 2017 | |
| Cuda code performance | 14 | 3149 | December 16, 2014 | |
| Confused about GPU vs CPU speed in multiplication | 8 | 6548 | February 19, 2009 | |
| GPU and CPU don't run in (pure) parallel ? | 24 | 20148 | May 4, 2007 | |
| GPU Perfomance How much GFlops??? | 27 | 37403 | August 30, 2009 | |
| Cuda program results are always zero in HW, correct in EMU? | 35 | 11161 | May 23, 2010 | |
| Wishlist Place your considered suggestions here | 201 | 204317 | April 13, 2009 | |
| Memory problem? ...incredible slowdown | 29 | 16301 | January 30, 2011 | |
| Is GPU worth it? GPU currently too slow. | 16 | 6040 | December 8, 2008 | |
| Problem with performance running parallel blocks | 3 | 582 | March 26, 2018 |