When Tuesday does the G70 NDA expire?

I read that Nvidia off loads vertex work to the CPU when the GPU is busy. Could this account for the relatively large differences in performance with faster CPUs?
 
g__day said:
Interesting look at 3dMark05 and how CPU bound the 7800 GTX is - especially in SLI mode.

http://www.vr-zone.com/?i=2367&s=12

I can verify this to the extreme... in my home system (Athlon64 3200+ @ 200x10=2GHz which means not overclocked, DFI Lanparty SLI-DR, 1GB RAM), a Geforce 7800 gives me 6777 3DMarks.. I am still investigating whether something's wrong with my setup, but I don't think I'll find anything... :eek:

edit: I was getting 6240 with a Radeon X850 XT overclocked to 540/580 btw.
 
egore said:
I read that Nvidia off loads vertex work to the CPU when the GPU is busy. Could this account for the relatively large differences in performance with faster CPUs?

Vertex FETCH limited.

[Mr. Ryszard's role played by geo, temporarily]
 
egore said:
I read that Nvidia off loads vertex work to the CPU when the GPU is busy. Could this account for the relatively large differences in performance with faster CPUs?

The graphic bellow the one with the CPU overclocking has the GPU overclocking and there is almost no diference between scores.
 
Simple question time: vertices are fetched via the CPU, the system bus (system RAM via AGP/PCIe), the vertex shaders, the card's RAM (via crossbar)? Given the CPU scaling, I'm guessing they're fetched, or at least produced, by the CPU?

I wonder what score a P4 system gives. :devilish:
 
It really depends upon how the application is built. There are ways to let the graphics card do nearly all of the work, but that's not always fastest.
 
Pete said:
Simple question time: vertices are fetched via the CPU, the system bus (system RAM via AGP/PCIe), the vertex shaders, the card's RAM (via crossbar)? Given the CPU scaling, I'm guessing they're fetched, or at least produced, by the CPU?

I wonder what score a P4 system gives. :devilish:

For the older hardware case, as far as I'm aware: the vertex buffers are placed all over card memory by the app, which makes it harder for the VFE to fetch efficiently than if they were in contiguous blocks. Not sure if there's scope for the buffers to actually fragment, but they're very large and often not placed near each other.

So it's the pull out of card memory to feed the vertex hardware where you're limited in 05, on most hardware.

With the 7800 SLI case it appears you're not completely VFE bound any more, you're more geometry production bound, getting the verts produced on the CPU before throwing them over the bus. The VFE unit did get tweaked for G70, and with SLI that's probably combining to shift where you're bound to somewhere else (at least partly).

Be nice to see 6800 SLI in the same tests, to confirm (at least G70 vs NV40).

Cheers geo!
 
Since its relevent to said discussion. Thought people might like to take alook at 7800GTX resolution scaling in 3dmark05. First person to guess which game test is actually fillrate bound wins a cookie! 8)

3dmarkscaler.png
 
JoshMST said:
Oooh, oooh, oooh! I know! Pick me! Pick me!

Considering you're obvious insider info. I'm afraid I'm going to have to exclude you. How do I know you havent been tipped off by insiders at ATI and Nvidia? :LOL:

wireframe said:
Is 3dMark05 Game Test 3 the forest one?

No Game Test 2 is the forest. Flying ship with all the water and the beast is game test 3.
 
Nice graph Chris. I would never have thought it.

Does anyone know if this this sort of geometry limitation is representive of modern games as well, or do they tend to be still fillrate limited ?
 
dizietsma said:
Does anyone know if this this sort of geometry limitation is representive of modern games as well, or do they tend to be still fillrate limited ?
It's not representative of modern games at all.
 
Back
Top