I looked at all the information that was leaked and I determined that the only missing pieces of information that I need to make a complete addition to the database were the Texture Unit count, specific Release Date and the Max Power Draw. Those missing details were not enough to stop me, so I made some educated guesses so that our readers can compare these cards to HD 7000 series and the rest of the cards in our database.
I won't go into how I estimated the Release Dates and Max Power Draw, but I will describe how I chose the Texture Unit (TU) counts that I did. There is always a relationship between the Stream Multiprocessors (SM), Stream Processors (SP) Texture Units (TU) and Raster Operators (ROPs). For example, the GeForce GTX 580 had 16 SMs. On each SM there was 32 SPs (16 x 32 = 512 total SPs), 4 TUs (16 x 4 = 64 total TUs), and 3 ROPs (16 x 3 = 48 total ROPs). So for every 4 TUs there were 3 ROPs.
According to the leaked information, the GTX 680 will have 32 SMs, 1,024 SPs and 64 ROPs. That means that each SM will contain 32 SPs and 2 ROPs. Now if the ration between TUs and ROPs for the GTX 680 is the same as the ratio for the GTX 580 (4:3) then that would mean that the GTX 680 would have 85.33333 TUs. 85 and 1/3 is not a round number so that means that the TU:ROP ratio for the GTX 680 will probably not be 4:3. So what will it be then?
How about 5:4? That would give us 80 TUs (which sounds like a good number) but then how would those 80 TUs be divided amongst the 32 SMs? 80 divided by 32 equals 2.5, which besides not being a whole number means that each SM would have two full texture units and then it would split a full one between itself and its neighboring SM. That does not sound very efficient to me.
How about 3:2? That would give us 96 TUs to be split amongst our 32 SMs. 96 divided by 32 equals 3. Well looky there boys! We got ourselves a whole number of TUs per SM! That sounds like a good candidate but if you look at the texture fill rates of a 96 TU GTX 680 and compare it with the HD 7950 the GTX 680 loses 81,600 to 89,600 MTexels/sec. I think the GTX 680 will probably have the same if not more texturing power than the HD 7950 so I think I need to try again.
How about 4:2 (or really 2:1)? That would give us 128 TUs (the same as the HD 7970!) and 128 divided by 32 is a nice even 4 TUs per SM. 128 texture units gives the GTX 680 a texture fillrate of 108,800 MTexels/sec which is more than the HD 7950 and equal to the HD 7970 when you overclock the GTX 680's core to 925 to match the HD 7970 core speed. That is why I think this might be the right amount of texture units.
Now that I have exhaustively explained my math, please let me know how many texture units you think will be on each stream Multiprocessor!