I'm curious if it will look like 2 chips or 1 to CUDA and if all 12 GB will be accessible to both of the chips or if it is split.
It could potential be much better for certain applications if it acts like 1 card and all 12GB of memory are allotted in a singe chunk of global memory.
I would buy one in a heartbeat instead of the tesla k40 I had been looking at.
I suspect that the Titan Z be like my GTX 590s (1.5 gigs per GPU, but marketed as 3 gigs) and GTX 690s (2 gigs per GPU, but marketed as 4 gigs) and could be really the GTX 790 in all but name [
http://videocardz.com/49465/asus-geforce-gtx-titan-black-6gb-memory-gets-listed-974-eur ], with 6 gigs per GPU. Applications like OctaneRender will probably see only 6 gigs. So the Tesla K40 with one GPU processor and 12 gigs of vram will load more data and still has a reason for existence, but the Titan Z will process that data about twice as fast (with double the CUDA cores) if the data doesn't require more than 6 gigs of memory space.
Also keep in mind, as Riggles first pointed out to me and as can be seen from the video, the Titan Z will consume 3 slot spaces
*/. At best, I could install only four of them and one Titan Black Edition (25,920 CUDA cores total) into my otherwise 8 double wide GPU slotted Tyan server for about $13k or I could install 8 Titan Black Edition SCs for about $8.08K (23,040 CUDA cores), saving me about $4.9K, but leaving me with 2,880 fewer CUDA cores. 23,040 CUDA cores happens to be the same number of total CUDA cores that I have at present in my 8 GTX 780 Ti ACXs that cost about $250 less per GPU card than the Titan Black Edition. The Titan Black Edition does have twice the amount of vram (6 gigs vs. 3 gigs), however.
*/ Note that in the video that Nvidia uses the price for 4 Titan Zs (4x$3K=$12k) whereas they appear to be comparing only three of them to attain the processing power of the Google brain system aggregation.