NVIDIA Clock Domains: Information now right in front of us.

[trinibwoy]

I assume your GPU is typically clocked at 450 mhz?

Yep.

 

[trinibwoy]

What I find strange is that there is no dual-display support for the performance monitor - I had to disable one monitor to get it to work. How do devs use the monitor while running their 3d apps?

 

[Geo]

AGP 6600GT stock clocks gives:

GPU_HOST = 277780 (no counters)
GPU_Graphics = 299250 (8 counters)
GPU_Raster = 299250 (0 counters)
GPU_Shader = 299250 (5 counters)

So the both the NV47 and NV43 share the same Host clock speed, which suggests (as the name sort of implies) that it is a reference clock within the core somewhere.

Oh. Right. The "2D mode" clock (i.e. GPU_HOST).

Still wondering what is left in "GPU_GRAPHICS" after cutting out "Raster" and "Shader".

 

[Arun]

What I find strange is that there is no dual-display support for the performance monitor - I had to disable one monitor to get it to work. How do devs use the monitor while running their 3d apps?

Windowed Mode FTW!
I'll admit I should stop being a cheapskate and buy a second monitor, though.
As for GPU_Host; are you sure it isn't the AGP/PCI-E bus frequency?

Uttar
P.S.: Stop arguing about the "counter" thing, the answer was in my original (edited a few mins after postage) message, damnit!

 

[Humus]

Too bad all of the counters appear with the value "0.0" for all the rendered frames.

A64 3000+
ATI X800
WinXP x64
Catalyst 5.7
DX SDK 2005 Feb

Hmm, that's odd. Try using the latest DX SDK.

Edit: Don't know really, but Win64 might also be a reason why it doesn't work.

 

[Geo]

Windowed Mode FTW!
As for GPU_Host; are you sure it isn't the AGP/PCI-E bus frequency?

277780 is suspiciously close to the 275000 I would expect for the 2D clock, and the only one of those in the appropriate neighborhood. And if that isn't it, then where is it?

 

[Neeyik]

The GPU_HOST can't be 2D mode for my 6600GT as it's 300MHz for that (as one can see by the 299250 for the other GPU counters). 277780 is quite a bit higher than 266666 to be the AGP speed though.

 

[Rys]

As Uttar speculated, GPU_HOST is more likely the timer source for other GPU functions like the PCI Express interface, etc

EDIT: Oh, and some cool 7800 GT clock domain info is already in this thread

 

[Geo]

Another year or so of practice at obliqueness and you´ll be ready to rule the universe at your father´s side, Rys.

Hrmph. I´ll play --GT has a 6.6% upclock too?

Edit: Hrmph! You and EAP!

 

[Arun]

As Uttar speculated, GPU_HOST is more likely the timer source for other GPU functions like the PCI Express interface, etc

Tsk tsk - it's only "speculation" if it's wrong, didn't you know that already? I'm curious as to whether that means there is a part of the chip clocked at the frequency though - or would that only be the frequency of the actual bus? The former seems more likely to me, but I could be wrong of course.

Uttar

 

[Geo]

Digit-Life in their GT review is claiming the upclocked domain is vertex units. Which I was pretty sure we'd ruled that out.

http://www.digit-life.com/articles2/video/g70-4.html

Edit for brainfart.

 

[Geo]

New RivaTuner out, including:

Now RivaTuner is able to monitor independent clock frequencies for core clock / geometric domain, core clock / shader domain and core clock / raster domain of G70 GPUs.

But I think I still don't get it. Have we determined that the geometric domain was the one being upclocked? If so, why didn't benchies show it?

 

[trinibwoy]

New RivaTuner out, including:



But I think I still don't get it. Have we determined that the geometric domain was the one being upclocked? If so, why didn't benchies show it?

What do you mean? 3dmark05's vertex tests show a difference for me between 480 and 499 - they result in clocks of 486/520 and 486/540 respectively.