Quick Question Regarding nVidia's Next Gen...

[Rroff]

Gasp a new upcoming chip is better than an older chip... oh the horror.

 

[An Exception]

Gasp a new upcoming chip is better than an older chip... oh the horror.

It is when they are on the same 40nm chip!

Usually I would only expect this kind of performance difference per mm2 if one chip had a process node advantage...

What it means is Kepler needs to be hugely more efficient than Fermi. i.e. it need to ditch GPGPU based architectures to survive, do you think Nvidia will do this?

 

[Rroff]

They won't ditch the GPGPU functionality.

I'm not an expert enough on the architecture to explain what I understand to a 3rd party - but the way the design was shoe horned down to 40nm sacrifices performance in several key areas, in theory on 28nm several areas can be run at full clock or even potentially given a good process double the shader clock ala Pentium 4.


(infact given the info on Kepler I wonder if this is what they intend to do).

 

[Lightnix]

They won't ditch the GPGPU functionality.

I'm not an expert enough on the architecture to explain what I understand to a 3rd party - but the way the design was shoe horned down to 40nm sacrifices performance in several key areas, in theory on 28nm several areas can be run at full clock or even potentially given a good process double the shader clock ala Pentium 4.


(infact given the info on Kepler I wonder if this is what they intend to do).

So, let me get this straight - what you're saying is what should happen is the 'un-shader' stuff should be running at the same speed as the shaders. And then, your suggestion is that they may once again double the clock speed of the shader domain over the rest of the core, essentially bringing us back to square one, under the pretence in that this would allow the whole functional part of the chip to run at in effect twice the clock speed it otherwise would do?

 

[Rroff]

Thats not what I'm suggesting. Texture management, polymorph and some other sub-systems are currently running half-clock some of which were originally designed to be running at full-clock "on the right process" - I've not followed the development close enough to know if they can regain that without significant re-design.

I am wondering given some of the info on kepler tho if they don't intend to implement a double (shader) clocked domain for some functions ala the Pentium 4 - tho that didn't end too well for the P4.