X-bit labs

From Extreme to Mainstream: Nvidia GeForce 8800 GT from Leadtek (page 4)

Category: Video

by Alexey Stepin , Yaroslav Lyssenko

[ 12/03/2007 | 03:09 PM ]

Raster Operators

The reduction of the number of raster operators shouldn’t be regarded as a bottleneck in the G92 architecture because the performance of modern GPUs with a unified architecture depends largely on the number and design of shader processors and on the efficiency of the driver’s shader code translator. The GPU’s 16 ROPs can only have a serious effect on performance if you use a very high resolution together with full-screen antialiasing, but the memory bandwidth is going to be a narrower bottleneck then. Moreover, the GeForce 8800 GT is not actually meant for such extreme operation modes which are a prerogative of the most expensive single cards and multi-GPU configurations.

The G92’s memory controller has 4 sections instead of 6 as in the previous chip and the total memory interface width is 256 bits now (64x4). That’s a recession in comparison with the GeForce 8800 GTS, but it is partially made up for by the increased memory frequency. Some improvements may have been introduced into the memory controller architecture to boost its efficiency. Don’t forget about the capacity: we know that 320 megabytes of memory on board the GeForce 8800 GTS 320MB is sometimes not enough in modern games and the GeForce 8800 GT, at least in the senior configuration, has an advantage here. 512 megabytes of graphics memory should be a positive factor under real-life gaming conditions.

Shader Processors

When it comes to shader processors, the GeForce 8800 GT is undoubtedly superior to the GeForce 8800 GTS. It not only has more ALUs but also has them clocked at a frequency which is higher than that of the GeForce 8800 GTX. Of course, this should have a positive effect on the card’s performance in modern games which make wide use of visual effects created by means of complex mathematical computations. The configuration of the shader processors hasn’t changed: 112 superscalar ALUs are grouped into 7 units with 16 ALUs in each (as mentioned above, there is probably a disabled eighth unit). One shader code branch can vary from 16 to 32 pixels like with the G80 chip.