RadeonSI Gallium3D Is Improving, But Still Long Shot From Catalyst

[log0]

I've heard that someone is working on a kcm module for kde 5 to control Mesa settings in the system settings framework. That was a while ago, though.

I find it odd that in that span of cards they all perform within pretty close margins of each other. I wonder how the 260x / 7770 perform? Are they in that same envelope? If that is the case, anyone looking at radeonSI gpus would get the best bang for the buck at the lowest end of the scale.

But that doesn't seem right. I thought the patches to enable all the cores went in a few Mesa versions ago? Why would hardware with only 2/3 the shaders (7850 vs 7950) perform on par or better in all these tests?

That is because they all have got 32 ROP units. So what this benchmarks are really testing is how fast this GPUs can push pixels onto screen. As others already stated benchmarks beyond 100fps are pretty much useless, you could simply use glxgears instead.

These cards need vertex and shader heavy food. The 7950 has a theoretical peak performance of 2.8 TFLOPS? Its compute units are just idling along with loads presented here.

[Adarion]

Good to see the progress. This IS pretty impressing. Yes, maybe fglrx is in most parts 2x the max. fps, but 120 fps is already a lot. I'm really impressed. If issues solve at this speed, then I hope Kabini/Kaveri is going to lift off soon on Linux. I can't wait for mainboard vendors to support Kaveri and bring Kabini boards and to see the 45W Kaveri in stores. Hopefully I can gift myself one for birthday this year.

[Andrecorreia]

the really problem is not 3d right now...

the really problem is not 3d right now, is the 2d performance, what i see in phoronix last 2d benchs, opensource are really bad

[zanny]

Wow, the open source driver has been making some amazing progress recently... actually pretty surprised. I can't wait until the performance is on par with the Catalyst driver, because it's plagued with bugs and terrible 2D support. All we need now is a control panel for Gallium3D to quickly tweak options like vsync and power management so it can be a viable replacement for ALL users.

Thanks for providing these benchmarks & keep up the good work!

I've heard that someone is working on a kcm module for kde 5 to control Mesa settings in the system settings framework. That was a while ago, though.

I find it odd that in that span of cards they all perform within pretty close margins of each other. I wonder how the 260x / 7770 perform? Are they in that same envelope? If that is the case, anyone looking at radeonSI gpus would get the best bang for the buck at the lowest end of the scale.

But that doesn't seem right. I thought the patches to enable all the cores went in a few Mesa versions ago? Why would hardware with only 2/3 the shaders (7850 vs 7950) perform on par or better in all these tests?

[smitty3268]

I found it interesting that the older games were at around 50% and the newer games were around 75%. Looks like one (or more) of the older gl commands is really slow.

I think the newer games just use more shaders, which are just running full speed on the gpu hardware. The older games are probably more bandwidth/cpu limited, where the si driver clearly is not as optimized.

[smitty3268]

Of course people care when catalyst runs 35% faster. Efficiency matters always. It's not like catalyst on Linux is on par with Windows to begin with.

What, power efficiency? Most people are running these games as fast as possible anyway, so it doesn't really matter whether you are running at 300fps or 350fps. They're both going to be maxing out the power requirements.

[smitty3268]

So what do you do if you have a lower end GPU that isn't capable of 60+ FPS because of horribly inefficient drivers? Of course you will care. There is a large margin of difference between Catalyst on Windows to Catalyst on Linux to the latest open source drivers. I'd rather see a day where the open source Linux drivers are more efficient than the Catalyst Windows drivers.

This is exactly the kind of misguided thinking that poor benchmarks like this encourage.

The limitations a driver has when it's running > 100fps have NOTHING to do with the limitations you might run into on another card at a slower speed.

NOTHING AT ALL. THERE IS NO CORRELATION.

It's highly likely the oss driver is still slower on those cards, but we have no way of knowing, and it's probably from completely different bottlenecks than these tests are showing here.


This is why people constantly point out that glxgears is not a benchmark. And everyone continues to use it anyway, seeming to think that higher glxgears scores must mean something on more realistic tests, even though people keep saying not to do so.

[Veerappan]

That's much closer to what I was expecting current RadeonSI to be capable of. Several months ago, the experience was acceptable for my system (7850 on a Phenom II x6) as long as I wasn't running multiple GPU-using programs simultaneously. Now it's doing great with all software upgraded to the latest code-bases.

There are still improvements to be made, but it's definitely usable from my perspective.

[brent]

You actually see these scaling effects in the benchmarks if you compare the 7850 results against the 7950 results. In many cases, both GPUs have almost similar results with open drivers, while Catalyst is able to push considerably more frames on the 7950 compared to the 7850. That very much looks like the open drivers are CPU limited.

[brent]

So what do you do if you have a lower end GPU that isn't capable of 60+ FPS because of horribly inefficient drivers

You shouldn't care too much about these benchmark numbers because performance doesn't scale linearly. radeonsi drivers have higher CPU overhead compared to Catalyst, which means they'll do worse in CPU bound situations. If you're pushing hundreds of FPS, you usually are CPU bound. Consider that when looking at the numbers. Lower-end GPUs should be closer to Catalyst.