LOSTCIRCUITS

Summary

One of the first companies to release a serial product based on the nVidia GeForce3 was MSI. The StarForce 822 follows the reference design for the GeForce3 but adds TV-in, TV-out (both analog and s-video) by means of the MS-5999 daughter board. The software bundle contains the MSI 3D-Turbo overclocking utility which allows settings of up to 350 / 580 MHz (core / memory), ULead video studio and a sampler of AVI files from Idea Design Center. The MSI Starforce 822 is extremely fast with the drawback that some applications run faster with older, less sophisticated cards, simply because they don't support advanced features. Some applications like Mercedes-Benz Truck Racing still don't work with the GeForce 3. This is a brief technical overview of some GeForce3 features leading into the specifics of the Starforce 822.

The GeForce 3 is the latest and most powerful graphics processing unit on the market. Manufactured at TSMC in Taiwan using 0.15 µm technology, the 57 million transistors only use some 4 Watt. Thus the power consumption is slightly above the numbers of the GTS 2 but acceptable enough, and shouldn't pose too much difficulty for most mainboards to cope with.

Memory Optimizations

Crossbar Memory Controller

One key feature of the GeForce3 GPU from the hardware point of view is the Crossbar memory controller. Standard graphics cards are using a memory interconnect to the GPU of 64, 128 or 256 bit fixed width and each frame to be rendered is completely transferred from the local frame buffer (graphics memory) to the engine. Very often, however, there are only minor changes from one frame to the next. If the entire information regarding the second frame is transferred again to the engine, the pixels that haven't changed contain redundant information that just waste bandwidth. The workaround is to split the memory bus in independently operating control units (in this case there are four individual controller of 32 bit width or 64 bit per cycle since the interface is DDR). Each of the individual controllers can access any part of the memory and move the data to the GPU. As a consequence, instead of having to transfer redundant information, the system can pick whatever new information is required to update the next frame whereas the redundant information is not being transferred. Because of the higher granularity and efficacy of this design, the crossbar memory controller makes much better use of the overall bandwidth available. Theoretically, splitting the controller into 4 individual subsystems could mean the equivalent of four times the bandwidth. A 1.5 x to 2 x scaling is more realistic, though.

Schematic overview of the Crossbar memory controller and its "associativity". Picture courtesy of nVidia.

Lossless Z-Compression

The z-buffer contains the depth coordinates for each pixel within a given frame. Any traditional graphics engine needs to read and write the z-coordinates for each pixel and, therefore the z-buffer traffic occupies probably the largest portion of the overall memory bandwidth available. The amount of data transfer can be reduced by using a compression algorithm, however, it is necessary, that no loss of data occurs, hence the term lossless compression. In the case of the GeForce3, the GPU uses a 4:1 lossless compression routine, which effectively cuts down the z-buffer traffic to ¼ of the original bandwidth required.

Next Page:    => Z-Occlusion Culling .... =>