Clock-For-Clock, Nouveau Can Compete With NVIDIA's Driver

[uid313]

Many games still don't work with nouvoue.
Corrupted, flicking textures, wrong colors, etc.

[birdie]

That's just excellent!

It means NVIDIA's GPUs are very well made so that you don't need to send them some special commands to get good performance from them.

[birdie]

Many games still don't work with nouvoue.
Corrupted, flicking textures, wrong colors, etc.

Writing GPU drivers is considered the most difficult and troublesome job in programming - it took NVIDIA years to polish their drivers and they have full specs for their hardware.

nouveau developers are basically banging against the wall trying to create a good open source alternative to proprietary drivers.

[pingufunkybeat]

The problem with Mesa drivers becomes much more obvious at high frames per second, because small inefficiencies in the driver are very large compared to the time spend by the GPU.

So it's pretty obvious that this will not scale. Mesa (Gallium) is efficient enough that it doesn't make a big difference at 10-20 fps, but when you render 300fps, it becomes very noticeable.

Is it possible to downclock Catalyst in a similar fashion and compare against that? The radeon tests are usually done the other way -- the open source is slowed down, and Catalyst run at full speed

[pingufunkybeat]

That's just excellent!

It means NVIDIA's GPUs are very well made so that you don't need to send them some special commands to get good performance from them.

This does sound like clueless blabbering, but there is some truth to that. Nvidia's older designs did put lot of stuff into hardware, which had to be handled by the driver on AMD cards.

The latest AMD generation (GCN) changed that, IIRC, so it's expected that the drivers will come closer to matching the maximum performance there.

[Rigaldo]

The problem with Mesa drivers becomes much more obvious at high frames per second, because small inefficiencies in the driver are very large compared to the time spend by the GPU.

So it's pretty obvious that this will not scale. Mesa (Gallium) is efficient enough that it doesn't make a big difference at 10-20 fps, but when you render 300fps, it becomes very noticeable.

Is it possible to downclock Catalyst in a similar fashion and compare against that? The radeon tests are usually done the other way -- the open source is slowed down, and Catalyst run at full speed

I don't think it has anything to do with clocks in the case of radeon .. But indeed, clocking down Catalyst would give radeon more beneficial results. But irrelevant.
I had normal clock speeds last time I used radeon.

[pingufunkybeat]

I don't think it has anything to do with clocks in the case of radeon .. But indeed, clocking down Catalyst would give radeon more beneficial results. But irrelevant.
I had normal clock speeds last time I used radeon.

Radeon runs at max clocks by default.

Still, it would be nice to see such a test. With all r600g bells and whistles turned on for a change.

[Rigaldo]

Radeon runs at max clocks by default.

Still, it would be nice to see such a test. With all r600g bells and whistles turned on for a change.

Oh, do you mean clocking down both Radeon driver and Catalyst to see how they compare then? Hmm ..
Hope I understood correctly this time.

[glisse]

This does sound like clueless blabbering, but there is some truth to that. Nvidia's older designs did put lot of stuff into hardware, which had to be handled by the driver on AMD cards.

The latest AMD generation (GCN) changed that, IIRC, so it's expected that the drivers will come closer to matching the maximum performance there.

There is virtually no difference in what is accelerated by hw or not since shader is the norm ie since dx10 GPU or since nv50 or r600 if you prefer.

Nor there is secret command on AMD GPU, anyone can go look at what fglrx is doing to see that for themself.

[pingufunkybeat]

Not GL functionality, it had something to do with command queue, or stream packing, or something like that, and might have had to do with the shader compiler and/or VLIW. Bridgman wrote about it, and thought that the Radeon/Catalyst gap should shrink as a result. Can't find the thread (here on phoronix) for the life of me now.