Reinventing the G80 - the NVIDIA GeForce 8800 GT 512MB


Reinventing the GeForce 8800

Reinventing the GeForce 8800

Moore's Law demands it, enthusiasts speculate about it and the head honchos at NVIDIA are probably rubbing their hands in anticipation at the prospects of world domination. In case you haven't guessed, the next evolution in the GeForce 8 series is set to be unleashed today and this new arrival has been predictably dubbed the GeForce 8800 GT.

There is an air of inevitability about this new high-end card. A different core from the ones found in the GeForce 8800 series (G92 instead of the G80) and a die shrink (NVIDIA's first 65nm core). Same old story, you may think and it won't be far off the mark, as it seems to be the tradition among chipmakers to produce a more efficient variant of their architecture. History also suggests that it is this second generation that will have the superior price performance ratio, which probably explains all the pre-release excitement surrounding this card.

The new numbering in the core hints that there have been significant changes in the G92 and browsing the documents released by NVIDIA, that seems to be the case. For one, despite the die shrink, the G92 core actually packs more transistors than the G80, at a whopping 754 million compared to the 681 million on the G80. It also exceeds the 700 million transistors on the 80nm ATI Radeon HD 2900 XT. What could be the reason behind the expanded core?

From a conceptual point of view, the G92 is a 'hybrid' (in an extremely loose sense of that word) of the G80 and the G84/86 cores. While the G80 undoubtedly retained superior performance in benchmarks, one feature it lacked compared to its lower end brethren is the presence of NVIDIA's second-generation video processor (VP2), which took the processing burden off the CPU when decoding high-definition video streams, leading to lower CPU utilization rates. Now this dedicated video processor has been embedded into the GeForce 8800 GT core and you'll get the full PureVideo HD deal, meaning that along with VP2, there's also the bitstream processor and the AES128 engine. This is a significant improvement over the older GeForce 8800 cards and the addition of this hardware is a major reason behind the transistor count boost. Yet another reason is the G80's supporting NVIO chip that took care of some of the display controller logic is also now embedded within the G92 core.

Of course, NVIDIA is not merely remaking a GeForce 8800 GTX Version 2 with the GeForce 8800 GT. The company is not replacing the G80-based cards yet. Hence, while the new card has the performance credentials of a GeForce 8800, it is slightly less well equipped than the reigning king of the GeForce 8800 series. This is reflected in the slightly lower number of stream processors on the GeForce 8800 GT (112 vs 128) and a lesser bunch of ROP units (16 vs 24) on the new card. However, the GeForce 8800 GT compensates with more aggressive clock speeds, with core and stream processor clocks running higher than the GeForce 8800 GTX while the memory on both cards are similar. If anything, it's the other high-end cards that should be more worried: the GeForce 8800 GTS looks to be eclipsed while ATI's Radeon HD 2900 XT may find itself knocked from its niche by the newcomer. Below is a comparison of the various high-end cards on the market with the debut of the GeForce 8800 GT.

Model NVIDIA GeForce 8800 GT 512MB NVIDIA GeForce 8800 GTX 768MB NVIDIA GeForce 8800 GTS 320/640MB ATI Radeon HD 2900 XT 512MB
Core Code G92 G80 G80 R600
Transistor Count 754 million 681 million 681 million 700 million
Manufacturing Process (microns) 0.065 0.09 0.09 0.08
Core Clock 600MHz 575MHz 500MHz 742MHz
Stream/Shader Processing Units 112 Stream Processors (operating at 1500MHz) 128 Stream Processors (operating at 1350MHz) 96 Stream Processors (operating at 1200MHz) 64 Shader units consisting of 320 Stream Processors (operating at 740MHz)
Texture Mapping Units (TMU) or Texture Filtering (TF) units 56 64 48 16
Render Backend Units (ROP) 16 24 20 16
Memory Clock 1800MHz DDR3 1800MHz DDR3 1600MHz DDR3 1650Hz DDR3
DDR Memory Bus 256-bit 384-bit 320-bit 512-bit
Memory Bandwidth 57.6GB/s 86.4GB/s 64.0GB/s 105.6GB/s
Ring Bus Memory Controller NIL NIL NIL 512-bit
PCI Express Interface Supports PCIe 2.0 (x16) x16 x16 x16
Molex Power Connectors Yes Yes (dual) Yes Yes (dual)
Multi GPU Technology Yes (SLI) Yes (SLI) Yes (SLI) Yes (CrossFire)
DVI Output Support 2 x Dual-Link 2 x Dual-Link 2 x Dual-Link 2 x Dual-Link
HDCP Support Yes Yes Yes Yes
Street Price US$199 (256MB) US$259 (512MB) ~ US$529 - US$549 US$279 - 349 ~ US$399