OpenGL Performance & Performance-Per-Watt For NVIDIA GPUs From The Past 10 Years

[phoronix]

Phoronix: OpenGL Performance & Performance-Per-Watt For NVIDIA GPUs From The Past 10 Years

Curious how the raw OpenGL performance and power efficiency has improved going back a decade to the GeForce 8 days? In this article is a 27-way graphics card comparison testing graphics cards from each generation going from the GeForce 8 series through the GeForce GTX 900 series and ending with the $999 GeForce GTX TITAN X. If you are interested in how graphics card performance has evolved, this is a fun must-read article.

http://www.phoronix.com/vr.php?view=22719

[Passso]

Thank you Michael for this impressive set of benchmarks!

I am always stunned by the way graphic cards evolve year by year, compared to CPU. The downside is that a 3 years old card is already... old.

[mannerov]

Hum, comparing with the recent open source radeon similar tests, it seems some ogl games have significant lead in term of performance on nvidia.

Difficult to know why.
Different renderer (because different ogl extensions available) ?
Some slow paths on mesa are fast paths on nvidia (and game would use these paths) ?
Nvidia driver multithreading does wonders for some use cases ? (likely for some of the cpu limited benchmarks).

The difference is clearly not just 'open source driver doesn't do full use of gpu features'. With these kind of differences, the gpu must end up underutilized because some slow paths are hit.

On synthetics benchmarks this seems better though (Valley is comparable according to the benchmarks, and by trying to reproduce Furmark test my tonga get similar to GTX 980).
I hope Vulkan solves these issues (and the nine state tracker does solve it already, but only for d3d9 games).

[birdie]

CS:GO results for GeForce 8xxx and 9xxx series are clearly wrong. It might be a regression or something but they are completely wrong.

[mannerov]

CS:GO results for GeForce 8xxx and 9xxx series are clearly wrong. It might be a regression or something but they are completely wrong.

The slow performance for these cards should be explained by the resolution used: 2560x1600. Likely they don't have enough memory to execute correctly the apps, and it has to remove/restore textures in memory all the time.

[schmidtbag]

I am always stunned by the way graphic cards evolve year by year, compared to CPU. The downside is that a 3 years old card is already... old.

It is unfortunate, though not all that surprising. The demand for GPU performance dramatically increases, while the demand CPU performance (for the average user) has pretty much reached it's peak 5 years ago. The only reason to get a new CPU is for energy efficiency, but GPUs still have a long way to go for performance.

[Michael]

Thank you Michael for this impressive set of benchmarks!

I am always stunned by the way graphic cards evolve year by year, compared to CPU. The downside is that a 3 years old card is already... old.

Thanks. Yep, time for upgrade

[librin.so.1]

CS:GO results for GeForce 8xxx and 9xxx series are clearly wrong. It might be a regression or something but they are completely wrong.

The benchmarks were done on the 361.18 drivers. And with that branch of these drivers, CSGO and some other OpenGL games have a major stuttering plus a performance regression issue. So, that might be the culprit.
And it's not just on Linux – people on reddit were reporting that upgrading to 361.xx drivers on Windows results in a major stutter + performance regression problem on CS:GO and thus quite some people on Windows are rolling back because of it.
(considering that the Linux blob and Windows version of the driver share a majority of the codebase, it's not surprising.)

[xorbe]

The slow performance for these cards should be explained by the resolution used: 2560x1600. Likely they don't have enough memory to execute correctly the apps, and it has to remove/restore textures in memory all the time.

This right here. The 256MB (!) / 512MB / 1GB cards were likely idled waiting for textures to swap in from system ram. Messes up both the power and relative fps readings, from a certain perspective. Need these "heavy" tests, plus some "light" tests.

[Jumbotron]

I am still amazed at how much the good 'ol GTX 680 is still a relative beast.