Radeon Gallium3D With Mesa 8.0: Goes Up & Down

[phoronix]

Phoronix: Radeon Gallium3D With Mesa 8.0: Goes Up & Down

Now that I've shared eight good features and eight bad traits of Mesa 8.0, which is the open-source graphics hardware library that's now supportive of OpenGL 3.0, it's time to begin looking closer at the performance. In this article are benchmarks of ATI/AMD Radeon graphics cards spanning five generations to show how the Mesa 8.0 performance compares to the previous release (Mesa 7.11) and the proprietary Catalyst driver.

http://www.phoronix.com/vr.php?view=16931

[Temar]

Hmm, Mesa is up to 10 times slower.

Why is the performance so miserable? Having modern OpenGL support is worth nothing if the performance is that bad.

[HokTar]

Jerome Glisse has just posted 2D tiling patches to the mailing list for r600g. That might help a bit.

[Pontostroy]

Why lightsmark is so slow it this test? With swapbufferswait=off end pcie_gen2=1 radeon hd6770 show 130 fps in 1920x1080.

[whizse]

Jerome Glisse has just posted 2D tiling patches to the mailing list for r600g. That might help a bit.

It should apparently give (up to) a 30% boost.

[whizse]

Hmm, Mesa is up to 10 times slower.

Why is the performance so miserable? Having modern OpenGL support is worth nothing if the performance is that bad.

People hack on difference things and have different priorities.

In this case the biggest push for GL3 is Intel and I think their driver is much closer to living up to the potential of the hardware. I also think they are better off when it comes to support for as-of-yet unreleased hardware (Ivybridge).

[drag]

Hmm, Mesa is up to 10 times slower.

I am being a bit of a dick here; but that sort of phrase never made any sense to me. How does that math work, exactly? How do you multiply a positive number by ten times to arrive at a lower number? I can understand that 'Mesa was one tenth as fast as Catalyst' or that 'Catalyst is ten times as fast as Mesa', but saying 'Mesa is ten times slower then catalyst' just seems really awkward.

Why is the performance so miserable?

I think that you are seeing the effect of having a lots of man hours being poured into application-specific optimizations. Lots of benchmarking over and over again and tweaking the driver to create lots of special cases and optimized code paths for popular game engines and common application developer patterns.

Having modern OpenGL support is worth nothing if the performance is that bad.

Because benchmarks do not tell the whole story. If I am building a computer for the sole purpose of rendering a game 4 times faster then my monitor can display it then there would be never be any point in caring past benchmarks. But the Catalyst drivers suck. Not only are they a pain to deal with and are closed source, but they are unstable, don't play the games that I use very well, the 2d performance tends to be poor... it can't even render my desktop correctly.

If you want decent OpenGL performance with a OSS system then ATI is your best bet.

[FourDMusic]

Michael, I'm not a journalist so I might be out of place, but could you please stop writing what your graphs clearly depict?

When running the open-source Nexuiz game, which is also one of the more demanding Linux games where Mesa is good enough to handle, there weren't any major performance changes here for the five ATI/AMD graphics cards on the R300g and R600g drivers between Mesa 7.11 and Mesa 8.0-devel.

My three year old daughter, which I don't have, could have told me this.

If there are words with the graph, then I expect them to describe additional detail, for example, why the open source driver performance under HD6950 in Nexuiz using Mesa 8.0 is near parity with Catalyst while for all the other cards, the Mesa 7.11 and 8.0 are approximately equal and about <75% Catalyst performance (page 5). I feel like you are wasting your time typing this superfluous information and definitely wasting my time when I skim through, trying to find the interesting bits of the article.

</rant>

[Luke_Wolf]

Hmm, Mesa is up to 10 times slower.

Why is the performance so miserable? Having modern OpenGL support is worth nothing if the performance is that bad.

Actually if you tabulate the provided data the opensource ATI drivers are running at an average of ~ 49% of the speed of the Catalyst drivers. which is really rather impressive given such a small team of people with drivers lacking features and optimization.

[Welsh Dwarf]

Michael, I'm not a journalist so I might be out of place, but could you please stop writing what your graphs clearly depict?

I'm not a journalist either, but in university (college for those across the pond) it's pretty much standard practice to comment graphs stating what you see, and only afterwords adding conclusions (this latter part being 90% of the work after the experiments are done).

Back on topic:

The good:
-Adding more demanding games to the list of benchmarks (adding glest and/or spring based benchmarks would also be appreciated).
-The double comparison of revisions and hardware

The bad:
-Not compiling in floating point textures and S3TC to be able to add Unigine to the list (No software patents in Europe)
-We could really do with your configure options and compiler flags to make your benchmarks more reproducible

The ugly:
-Not stating that in FOSS, first you make it work, then you make it work well and finally you make it work fast, and that the last 6 months were about parts 1 and 2.