Intel Linux Graphics Performance Q4'08
Phoronix: Intel Linux Graphics Performance Q4'08
The past year has brought several invasive changes to the Intel Linux graphics stack with the introduction of the Graphics Execution Manager for GPU memory management within the kernel, support for the Direct Rendering Infrastructure 2, and kernel mode-setting finally getting ready to enter the limelight. How though has the work this year affected the overall performance of Intel integrated graphics on Linux? In this article we have run a few benchmarks that show where the driver was at a few months ago and where it is today.
Hmm, this was very horrible to read, there must be something very wrong here, well its a bit left before release
Ouch! Can something like that be posted so close to Xmas?
Damn, a friend is looking to buy a 4500-based laptop, but with this kind of performance it will be unusable.
Michael, can you comment on regular computing tasks? Can the driver play fullscreen video? Is compiz usable (e.g. when browsing the web)?
i had to do LIBGL_ALWAYS_INDIRECT=1 before running any glx app to get good performance with uxa and dri2.
Without it glxinfo only reports software rasterization.
It's OK, the open-source coders will save the day of course. Don't worry about it, they got it covered!
That test is wrong. You don't compare different drivers on different distros. You compare them on the same distro. Now I get to say that it's not the driver that is slower, but it's the newer Ubuntu version that slows it down.
Kernel, xorg-server, mesa, mesa dri drivers and intel drivers are too many variables to compare if you want meaningful results. Moreover, distros adapt certain settings (like whether to choose exa or xaa, whether to enable XAANoOffscreenPixmaps, etc.) during the course of their beta cycles. Using a beta distro may therefore not tell all that much. What you should do is try to single out where the problem lies by using the same intel driver with different xorg-servers and mesas and then different servers with different drivers, etc.
My experience is that the exa acceleration method on intel regressed slightly from the i810 to the intel driver and incredibly much from X 7.3 to 7.4. I have thus switched back to xaa. With XAANoOffscreenPixmaps and EnablePageFlip it renders slower in recent xorg-servers (including 18.104.22.168), but is still acceptable for compositing window managers. I suspect regressions in mesa as well, but then git snapshots do not tell all that much. The new mesa is supposed to be released this month btw.
Much in X is in a transition period that has started a few months back and will come to a close in early 2009. Most of those exciting things you read about on phoronix will be delivered for intel cards first, thanks to openness and good support! For now, you can use an older X (no one forces you to use Ubuntu alpha releases) and simply wait until the dust has settled.
Originally Posted by BlackStar
But even if you go for unstable stuff, in my experience, all recent snapshots of mesa, xorg, intel drivers, etc. are good enough for compositing (i.e. compiz or kwin) if you play a little with xorg.conf. Moreover, your favorite distro will probably avoid the more problematic releases.
I find it pointless to publish benchmarking results from the Ubuntu development version just a few days after its introduction of a new major version of the xserver and other bits. At this point, you're happy if the Xorg stack builds and runs. Now comes a period of ironing out compatibility issues and other newly discovered bugs before performance can be optimized.
To compare intel 2.4 vs 2.5 I can suggest the intel 2.5 and corresponding libdrm packages for Ubuntu 8.10 in the "intel-gfx-testing" PPA (see the ubuntu-x mailing list for details). These packages have not been thoroughly tested either, but at least the Xorg base will be the same so there is a better chance the performance difference is actually due to the driver.