NVIDIA's Technology: G80For many years we have been dealing with separate pixel and vertex calculation engines, but with the come of Microsoft's new direct 3D api - DirectX 10 - things are about to change dramatically. The new approach is all about efficiency, ie. make better use of transistors so that less power is just flowing away without being used. For example, depending on in what scene you are, the software could ask for way more pixel calculations compared to vertex calculations. If your GPU had 24 pixel shaders and 8 vertex shaders, that would be no problem, all shaders would be filled and so the VGA card would be rendering at almost full power. But what if we come in another scene where the PC software requires lots of vertex calculations? Then the pixel shaders would be doing some calculations, but it might happen that they are only used for 50%, and thus the other 50% calculation power is just lost. Now... if we could somehow trick the pixel shaders to do some vertex shading operations, then wouldn't we pull up to efficiency up to 100% again?
The answer is yes, but the trick is not as simple is just saying: "hey, from now on you are doing vertex calculations in stead of pixel calculations". NVIDIA designed a new GPU existing out of stream processors. Each stream unit is a small processor which is completely programmable to do either pixel calculations, vertex shading, geometry shading and even physics calculations! So, no matter what needs to be drawn on the screen, in theory, the GPU will always be using 100% of its capabilities.
All this also means we will be no longer talking about pixel pipelines and vertex shaders.
The Geforce 8800 GTS has 96 stream processors. If you compare that to the 24 pixel shader and 8 vertex shaders we had with NIVIA's previous high-end card, the 7900GTX, we are talking about a potentially huge amount of rendering power. The addition of geometry and physics shaders will only improve image quality and average frame rate. The G80 GPU is build out of 681 million transistors. That's almost double the amount of the previous generation's G70 GPU. Since the G80 is still using the older 90nm process, the core is large and needs enough cool: a big heatsink.
| 7900GTX | 8800GTS | 8800GTX |
Core | G71 | G80 | G80 |
Productionprocess (nm) | 90 | 90 | 90 |
Number of transistors (million) | 302 | 681 | 681 |
Shaders | 24 pixel / 8 vertex | 96 | 128 |
ROP's | 16 | 20 | 24 |
Core frequency (MHz) | 650 | 500 | 575 |
Vertex frequency (MHz) | 650 | 1200 | 1350 |
Memory size (Mb) | 512 | 640 (320) | 768 |
Memory interface (bit) | 256 | 320 | 384 |
Memory frequency | 800 | 800 | 900 |
RAMDAC's (MHz) | 400 | 400 | 400 |
SLI compatible | yes | yes | yes |
PCI-Express | yes | yes | yes |
Both our Sparkle 8800GTS vga cards make use of Hynix HY5RS GDDR3 memory. This type of SDRAM is made for 1,8V supply voltage and should do its work up to 900MHz (1800MHz effective). Because of the late launch date of the 320MB 8800GTS, Sparkle equipped their cards with newer revision memory chips compared to the ones used on the 640Mb model. Hopefully this provides the tweakers out there with more overclocking headroom.
A third chip we found on our G80 board is being called the NVIO. This chip is mainly responsible for any in-/output that is not going through the PCI-E bus, for example Dual DVI, SLI, ...
I would like even more graphs of the AA tests because that's what matters not the number.