Workstation motherboards provide 4 or more PCIE x16 gen 3 slots with a good distribution of dedicated links direct to multiple onboard CPUs. With dual cards such as 7990 which are in fact 2×7970 each one can in theory load up a 4 slot motherboard with 8 gpus in total. Alas, in practice, there can be obstacles booting with 7 or 8 of them with certain motherboads as they tend to run out of PCIE resources after around 6 cards which could be considered to be surprising for a workstation class motherboard. Incidentally, I think it’s worth noting that not only are these the only model of 7990 made but they are practically impossible to get hold of so very rare gems indeed.
Watercooling four dual gpus on a standard workstation motherboard can also be a challenge due to a severe shortage of space between them. Dual slot spacing between PCIE x16 slots is a tight fit as the tubes can take up almost three slots worth of space between cards. In this configuration they are installed on alternate slots so if you had 7 slots in total you’d only install on 1, 3, 5 and 7 leaving 2, 4 and 6 empty. Though 2, 4 and 6 will usually be PCIE x8 slots anyway as opposed to the rest being PCIE x16.
As you can see below these cards have been completely stripped down of their air cooling and heatsink apparatus prior to attaching waterblocks and tubing for coolant to pass through them. The tubing is secured using barb fittings which are stronger than the alternative: compression fittings though they do lack the aesthetic appeal of compression fittings. Compression fittings can come apart under tension and that can create a real mess as I realised the hard way one night.
If not using full card waterblocks (which these aren’t) individual adhesive heatsinks for all the ram chips (known as ram sinks) are required for sufficient cooling. There may be upto 12 of these tiny little ram sinks on each face of a card. I don’t have any photos of that right now but I’ll try and get some. Though, ram sinks, can be flimsy and easily become dislodged and fall off the cards if they are knocked which is why some people prefer full cover blocks. Full cover blocks are more robust but also more expensive.
In terms of power the system is supplied with 2400 watts of power composed of two 1200W power supplies chained by an adapter for the first to kickstart the other on boot. Half the gpus are powered by one and half by the other. This particular machine has 128GB of RAM and dual xeons with a combined total of 32 cores. Update: As rightly pointed out by the commenter below I meant hardware threads not physical cores here.
Note: I do not own the hardware or the photographs but consent has been acquired to publish them here. Also this system is not conceived or assembled by me though I do have the pleasure of loading it with OpenCL benchmarks and computations.