Size and Power Requirements

The G80 cards can be big! The high-end ones (8800 GTS,GTX and ultra) are double-width, meaning you have to have allow two slots of space on the motherboard to accommodate each card. Each card only has one connector; the extra space is for the heatsink and exhaust fan. The cards are also relatively long. The GTS is shorter than GTX/ultra.

These cards also require relatively large amounts of power. The PCIe bus supplies some power, but supplemental 6-pin PCIe power connectors are needed for the high-end cards, one for the GTS and two for the GTX/ultra. High end power supplies are thus recommended, particularly for multi-GPU solutions. For tallying up a suitable power output, note that in principle each PCIe slot and each supplemental power connector can supply 75W. Modern power supplies often have one or more of these 6-pin connectors on them; graphics cards typically also come with adapters that convert two IDE hard drive power connectors to a single 6-pin PCIe one.

Bearing in mind the above-mentioned space and power requirements, I decided to get two 8800 GTS cards in the end. They both easily fit into the Sun Ultra 40 M2 host machine; see the picture below. Incidentally, this machine is wonderfully made and is very suitable for multi-gpu hosting; it comes with two 6-pin supplementary power connectors and a 1kW power supply!

I have also tried running one of the cards in my home machine, a cheap home-built system based around a Biostar T6100 motherboard and AMD64 processor. Luckily for me it happens to have an Nvidia chipset with integrated graphics and so works perfectly as a low-cost development machine. The no-name 450W power supply has so far been okay.

Note: NVIDIA have recently (Nov 2007) released a new mid/high end card, the 8800 GT, that is a standard single-width card needing only one supplemental power connector. It is not supported in CUDA 1.0, but is supported in CUDA 1.1. Then this could be an ideal card for CUDA for development, or, for the brave, for using multiple cards at once. They have also (Dec 2007) released a souped-up double-width version, called the 8800GTS 512Mb, also supported in CUDA 1.1. This latter card is in fact more like the 8800GT than the original GTS's. It also significantly outperforms the original GTS's and is probably on a par withe GTX; the main difference being the memory size and bandwidth.