NVidia's 'G80' GPU is built on TSMC's 90 nanometer
manufacturing process, and it contains a nice round 681 million transistors. If
you recall, the Geforce 7800GTX 'G70' had 302 million transistors, and by way of
comparison the ATI Radeon X1900XTX has 384 million transistors. In its current
form the G80 demands a lot of power, and generates a lot of heat; upwards of 80W-85W under load. This is why the reference videocard
design utilizes such a large heatpipe cooled thermal solution.
nVIDIA's 'G80' is the first DirectX 10 compatible GPU-series on the market, and it
is completely different from videocards of the past. The 'G80' utilizes a
unified architecture, which is another way of saying that it merges vertex and
pixel shaders into one floating point processor. There have been other GPUs to
take this route, notably the ATi graphics processor inside the Xbox 360, but
nVIDIA is first to release this type of technology for the home PC.
The GeForce 8800GTS is based on the same G80 core that powers the GeForce 8800GTX, however nVIDIA disables a few rendering pipelines here and there, and trims the memory bus to ensure that the GeForce 8800GTS is slower than the GTX. It's not as bad as it sounds, the GeForce 8800GTS 320MB videocard is still a wickedly fast platforms for games.
GeForce 8800GTX Core Shown
The nVIDIA GeForce 8800GTS GPU has access to six "Thread
Processors" as opposed to eight with the GeForce 8800GTX. Each has 16 Stream
Processors (SP), for a total 96.
Gone are the hard coded Vertex and pixel shaders, they have been
replaced with the more flexible Stream Processor that calculates both types of
data. The Stream Processors run at a blistering 1.35 GHz. Traditional core clock
speeds as we know it are dead, as several internal processors are running at
Each Thread Processor has two groups of eight SP, and
each group talks to an exclusive texture address filter unit as well as well as
being connected to the shared L1 cache. When more memory is needed the Thread
Processor connects to the crossbar memory controller. nVIDIA's crossbar memory
controller is broken up into five 64 bit chunks (six with the GeForce 8800GTX),
which means the width is essentially 320 bits wide (384 bits for the
the GPU towards a threaded design, the nvidia G80 is much more like a processor
than any graphics cores of the past. Any type of data, be it pixel, vertex, or
geometry shader can be processed within the SP. This allows load balancing to occur between the various tasks.
Although DirectX 9 does not support unified shader
instructions, load balancing is handled automatically by the GPU so it's not
something developers have to worry about. Load balancing ensures optimal
performance from the GPU no matter what the situation, as pixel and vertex
processing shifts greatly during game play.
While there are 96 Stream Processors broken down into 16
8-section chunks, there are only 32 texturing units total, 4 per thread
processor. The texturing units run at core clock speed, and as indicated can
handle 32 textures per clock. One of the most important "features" is that the
texture unit operates independently of the SP, so texturing can occur at the
same time as shader rendering.
One other new feature the G80 brings to the table is a
stand alone video display engine completely designed for this GPU generation.
Set off to the left of the G80 GPU is nVIDIA's new discrete display chip. This
chip supports both TDMS logic for LCD monitors and RAMDACs for analog displays.
By moving the display engine outside the GPU, it allows nVIDIA to have less
overhead with multi-GPU videocards.