NVIDIA 367.27 vs. 367.35 Linux Driver Benchmarks With GTX 1060/1070/1080

[phoronix]

Phoronix: NVIDIA 367.27 vs. 367.35 Linux Driver Benchmarks With GTX 1060/1070/1080

Released two weeks ago was the NVIDIA 367.35 Linux driver as the latest stable binary driver for NVIDIA hardware. Here are some performance tests to see if it upped the NVIDIA Pascal Linux performance at all...

http://www.phoronix.com/scan.php?page=news_item&px=NVIDIA-367.35-Benchmarks

[sdack]

Benchmarks are usually a good method for finding differences, but this result yields no explanation for why the driver was changed. Would be nice to know why Nvidia did it seeing how for some it is actually turned into a regression.

[LinAGKar]

That's some weird looking graphs.

[mike4]

not displayed correctly on Firefox/Ubuntu.

[bug77]

not displayed correctly on Firefox/Ubuntu.

Firefox/Kubuntu looking good over here.

[Hi-Angel]

not displayed correctly on Firefox/Ubuntu.

Disable adblock, or buy "premium".

[andre30correia]

it's only a bug fix release

[pszilard]

Hint: If you want to see interesting differences, I'd recommend including in the comparison the very _first_ driver released (367.18) and running memory-bound benchmarks (the Triad and FFT benchmarks should be suitable).

[sdack]

it's only a bug fix release

Not quite so. Most notably is the line saying, "Improved buffer write performance of the nvidia-drm DRM KMS driver by using write-combined DRM Dumb Buffers where available." This indicated to many careful readers a more fundamental change in parts of the driver. Nvidia, being in the business of writing x86-drivers for a long time, and any good software developer can tell you this, knows fundamental changes can bring many problems with them. These are often hidden problems that then can come to surface and that often only affect a minor group of people, but can be substantial enough to hurt the reputation. Therefore are changes like these not taken lightly by developers and require a good reason to be implemented. The reason is given in the release note ("improved buffer write performance"). Yet nothing seems to show and so it remains a minor summer mystery.