Well, I think we all want to know how this card really works beyond what NVIDIA has shared, anyway. We know we have a core clock and a shader clock. We also know there is the premise of a ROP buffer partition in the memory of the card (128 Mb for GTS and 256 MB for GTX). There also appears to be two unified shader "partitions" that are each unified which NVIDIA refers to as stream processors. So, instead of passing the work back to a common thread arbiter, it moves forward to a dedicated shader partition after the raster operation which was the driver for the hybrid rumors. It all comes down to math, since circuits need to base their operation on some sort of root oscillation.
What info do we have:
.............Core.......Shader
GTS........525........1200
GTX........575........1350
So, just doing some quick relation calcs between the two cards, we have:
Core * 3 less 375 = shader on both cards
What does all this mean? Who knows. But, we need to figure out the relationship between shader and core values as the core clock increases.
Does the shader makes "bumps" up the ladder? Does it stay flat?
What else do we know?
SM 3.0 performance is much better than SM 2.0, which indicates that the shader capacity is much better. Even more so than the vertex/shader ratio on a X1950XTX (thanks to the unified architecture).
The GTS appears to get more performance than the GTX clock for clock and is more easily overclocked. Why is this? Does the shader clock stay locked so that is never exceeded? Are the ratios somehow different than the GTX? Is the core partitioned into two with one being a lower clocked portion and the other higher clocked?
There is NO 2D performance. Since this is DirectX 10 and this will also be DirectX 9.0C driven for Aero, 2D clocks seem to be a waste which explains the higher idle temps. Last gen had a lower clocked, lower power 2D mode that subsequently reduced idle temps. The 8800 series run in 3D mode and voltage levels at all times.
So, share your thoughts, ideas and musings to see if we can pick this card apart. Doing so will help us wring out the last bit of performance.
Bookmarks