A Big Comparison Of The AMD Catalyst, Mesa & Gallium3D Drive

[cb88]

Yah.. r300g is still CPU limited at least to some extent. I'm only getting 10 or less FPS out of openarena on a 2xPII@300Mhz w/ radeon 9800 SE (not nearly as fast as the 9800 Pro but still respectible)

Course its common knowledge that openarena is quite CPU limited itself...

It might have improved some in the last month though.. since I checked

Are there any benchmarks/games that are severly GPU limited but light on the CPU I could try?

[89c51]

i read about CPU being the bottleneck (or one of) and i have to ask.

how come the blobs are not affected by it?? or is just that the other optimizations they have compensate for that??

[RealNC]

i read about CPU being the bottleneck (or one of) and i have to ask.

how come the blobs are not affected by it??

That's a wrong question. "It" doesn't exist as an entity. A bottleneck is not an entity that exists, it's just a behavior resulting from non-optimal code. Otherwise we would use bottleneck firewalls so that those evil bottlenecks are kept out :-P

The blobs are written with performance in mind. If they weren't, people would not buy the cards. The open source drivers and surrounding stack are not written with performance in mind. The developers struggle to make the stuff work as it is. Trying to match the performance of the blobs at the same time would probably mean that we would have to wait another 10 years for the drivers.

[smitty3268]

i read about CPU being the bottleneck (or one of) and i have to ask.

how come the blobs are not affected by it?? or is just that the other optimizations they have compensate for that??

To clarify, the blobs are CPU limited on most of those tests as well. It's just that they are optimized to use less CPU and so perform much better. I suspect that in a shader heavy test the results might be closer together because it would rely more on the GPU and less on the CPU, but I really don't know if that's true or not. Michael's never really tried that kind of situation on any of these tests.

[bridgman]

The binary drivers are also multi-threaded, which helps to hide CPU overhead *if* you have at least 2 cores.

Multithreading mostly helps when drawing is the bottleneck, however, but some of the initial profiling efforts suggest that state changes may be a bigger time sink than drawing right now. Not sure if that has been confirmed though.

[elanthis]

No, that won't cut it. Compiz *can* do what you suggest, but this 'feature' is extremely userunfriendly and is thankfully disabled by default. If you disable compositing for fullscreen windows and a notification pops up, you are in for a flicker-fest.

How about disable compositing only for full-screen windows that are above all other windows?

Enable compositing if something else pops up on top, full speed if there's nothing in the way.

I'm not familiar enough with how the compositing is enabled/disabled for existing windows and whether or not this is feasible.

[Kano]

With the exception of Q nobody would use a highend gfx card with oss drivers to play games. Some results are a bit weird that some tests are faster with a higher res but basically every noob should see that oss drivers are not made for gaming, they handle compiz, maybe some very old games but thats all.

[RealNC]

How about disable compositing only for full-screen windows that are above all other windows?

Enable compositing if something else pops up on top, full speed if there's nothing in the way.

Uhm, that's exactly how this is currently done in both Compiz as well as KDE.

[BlackStar]

Uhm, that's exactly how this is currently done in both Compiz as well as KDE.

Except when it isn't. It works that way in theory, yes.

[marek]

With the exception of Q nobody would use a highend gfx card with oss drivers to play games. Some results are a bit weird that some tests are faster with a higher res but basically every noob should see that oss drivers are not made for gaming, they handle compiz, maybe some very old games but thats all.

Well it really comes down to whether a game is both CPU-limited and unplayble on oss drivers rather than whether it's just a game. Some relatively new games do work very well. The article only tests free games, so the only thing we know is that free games are unoptimized with regard to CPU usage, but that's not a big surprise as their developers are understaffed as well. Most of the games tested in the article are ioquake3-based anyway, so it's like saying "hey you suck in ioquake3 and a couple of other benchmarks".

There is no commercial game tested in the article and don't say there aren't any...