I’ve been running CUBLAS matrix multiply on the TK1, and I’ve been observing some strange behaviors.
- When running the default size of 320x640, the performance varies from ~2.5-4.5 ms.
- When I turn off the desktop and just ssh in to run matrix multiply, the performance actually drops by almost 8x.
Desktop on matrix multiply (640x1280) ~10 ms
Desktop off matrix multiply (640x1280) ~85 ms.
Anybody have any idea why this might be happening?