I'm testing an Asus UX430UQ under Ubuntu. It has an NVidia 940MX graphics chip in it. I'm using Lightworks to benchmark performance, and the numbers are coming up heavily in favour of the open source
nouveau drivers over the proprietary ones in 4 of the 5 benchmarks.
I ran a bunch of tests with each of the drivers enabled (rebooting after switching the driver), discarded outlier values, and averaged the results of each benchmark for each driver.
Nouveau host to GPU: 998.1fps GPU to host: 516.4fps shader: 89024.87fps playback: 342.24fps render: 192.22fps Proprietary host to GPU: 208.41fps GPU to host: 251.66fps shader: 142293.05fps playback: 181.75fps render: 37.12fps
So other than shader performance at 60-70% of the proprietary result, the open source drivers appear to outperform the proprietary drivers by 2-5 times.
I've also found that real performance during playback of a particular section of an edit is noticeably worse with the proprietary drivers: there are some occasional dropped frames and audio glitches like I experience when playing the same section of that edit on my 2012 MacBook Pro.
Is there a reason why the open source drivers would be overall faster than the proprietary ones? Are these benchmarks misleading in some way? Have others seen similar results? Should I just use the open source drivers?
EDIT: Based on @Ken's suggestion, I checked the PRIME profile in
NVIDIA X Server Settings. It had been set to NVIDIA. Switching it to Intel (and logging out and back in) and running the Lightworks GPU benchmarks again gives results that closely match the open source result, suggesting that the
nouveau drivers were simply using the integrated graphics:
host to gpu: 969.2525 gpu to host: 516.655 shader: 87600.006667 playback: 337.516 render: 191.4125
I suppose some of these results make sense — host to GPU and vice versa, at least, should be fast since the chip is integrated into the CPU; better shader performance from the NVidia chip is to be expected. I would have thought that dedicated graphics would be faster for the render test, though.