Mesa 9.1 Results Are Mixed For Radeon Gallium3D

[pingufunkybeat]

which for a passive cooled card is bad news.

I have had two passive HD4000 cards in a multi-seat configuration and they've done their fair share of full load over several years, with no glitches.

The only problem is heat and electricity use, which is why I switch to low profile whenever I can.

[Hamish Wilson]

Thanks for the answer. I'm happy I won't have to change any config (even if it is simple)... now if swapbuffers was disabled by default I could finally drop xorg.conf and the problems I have with it

That is probably never going to happen to me because of my need for Zaphod Heads and the annoying fact that I need to enable Crtl-Alt-Backspace to actually work on modern X.org servers now. Why on earth do people keep turning it off?

[bridgman]

AFAIK newer cards go into a fairly low clock state by default. Transition happened mostly during the NI series IIRC.

[ua=42]

I'm a developer first, and I care a lot about cpu performance. Most of the things I do can also take full advantage of my six cores. Phenom X6 and a HD4350.

My primary criteria for a gpu: low tdp (<60W, electricity is not cheap), passive (no noise), runs on free drivers with most features.


That said, I do play some games, and also develop some 3d games and visualizations. For those I do care about these benchmarks. It just so happens that the other things have more priority for me than 3d oomph, but I can still care about getting more performance from my existing card.

Then you will want intel. Free drivers with Passive cooling would be a bad idea with AMD because it doesn't do any power saving at all (by default they have it go into its highest clock state and power use) which for a passive cooled card is bad news. I don't know how good the nvidia open source driver is on reclocking.... so you would have to ask someone else.

[stalkerg]

Strange!

Realy strange! I have HD6770 and with kernel 3.8, mesa from git I can get 40-60fps in OilRush with ULTRA QUALITY and 1920x1080!

Maybe hd6xxx have 4 async dma ring... hd4xxx have only one.

[Bitiquinho]

mesa >= 9.1 + kernel >= 3.8 -> hyperz enabled

Thanks for the answer. I'm happy I won't have to change any config (even if it is simple)... now if swapbuffers was disabled by default I could finally drop xorg.conf and the problems I have with it

[whitecat]

Right. However, high CPU usage by driver is a problem as well. And btw, opensource radeon driver is quite troublesome at this. I.e. there is severe CPU load when running 3D games. Attempt to profile this situation lead me to think that it's GPU driver who eats most CPU in it's ioctl. Unfortunately, tracing, profiling and so on has proven to be very delicate when it comes to graphics. I've got a number of hard lockups when trying to profile or trace things anyhow better so I only managed to get impression that ioctl of radeon's driver is a CPU hog. But not more than that.

The fact is that you do not understand what is the purpose of this benchmark. Michael was not benchmarking the performance of the 4650, nor the CPU usage... The topic was to see the difference of performance between differents versions of Mesa.

Moreover, a 4650 should handle all this games at hight rate FPS...

[curaga]

This is very unbalanced setup. It basically shows that you don't really care about 3D performance at all. So, do you really care about 3D benchmarks when running on this setup? Why? I fail to understand that.

I'm a developer first, and I care a lot about cpu performance. Most of the things I do can also take full advantage of my six cores. Phenom X6 and a HD4350.

My primary criteria for a gpu: low tdp (<60W, electricity is not cheap), passive (no noise), runs on free drivers with most features.


That said, I do play some games, and also develop some 3d games and visualizations. For those I do care about these benchmarks. It just so happens that the other things have more priority for me than 3d oomph, but I can still care about getting more performance from my existing card.

[tomato]

Reason for introducing other measurements than FPS. Since player on FLOSS drivers will care more about maximal delay between current and next frame.

Exactly, stuttering is much more irritating than running at a solid 20-30 fps.

[przemoli]

I was investigating why World Of Padman appears to be slower with Mesa 9.1.

The fact is it's not slower at all. The problem is the GPU hardlocks twice during the benchmark and the recovery takes 10 seconds, so in total the computer is frozen (0 fps) for 20 seconds. The cause of the hardlocks is Hyper-Z and the reproducibility of the hardlocks is 100%.

Reason for introducing other measurements than FPS. Since player on FLOSS drivers will care more about maximal delay between current and next frame.