Ever since the first Kepler (Nvidia 600 series) cards started landing on shelves and disappearing off of them into enthusiast systems, there have been indications Nvidia are preparing a bigger, in fact a huge-manatee sized card. Atomic noted about a month ago that Nvidia was to hold a seminar at their upcoming 'GPU Technology Conference' (GTC) discussing a "new 7-billion transistor GPU". Held today, the seminar unveiled what they’ve been working on: the Tesla 'High Performance Computing' server card.
The GK-110 chip powering Nvidia's K20 is simply the most complicated piece of machinery humanity has ever manufactured and if Mayans (people who died out due to forgetting how to grow corn) are right, it's the biggest we ever will manufacture. Packing in three times the number of component transistors of Intel's grandest server parts, GK-110 has space for a boggling 2880 CUDA cores, alongside a 384-bit GDDR5 memory bus that connects 6-24 GB of GDDR5 memory. The K20, alongside its GTX-690 based K10 sibling, is designed to function as a building block for future supercomputers - already being pegged for such projects as Titan (alone expected to use 18,000 cards) and Blue Waters. Interestingly in some areas performance is bested by AMDs 'Southern Islands' architecture, however for most tasks it does shine through as a true powerhouse.
In a move even closer to most Atomicans hearts than weather-predicting supercomputers, Nvidia have also quietly mentioned that K20 is "highly virtualisable" - it is designed to render graphics and deliver them remotely via the web. We'd suggest K20 and K10 are clearly intended to be the hardware powering the cloud-gaming 'Gakai/GeForce Grid' that David examined (also announced at the conference) . Cloud-gaming is an important future direction for graphics, however GK-110 should, as promised, also power a truly epic GeForce graphics card before the year is out.
The backstory is this: Nvidia's engineers have whispered of changes to their decade-long gameplan of sketching out massive GPU designs and then whittling them down to the midrange. Part obsession, part design philosophy, their goal has always been to create the most powerful card first, followed by several scaled-down versions - a 'top down' philosophy. In part driven by the sheer complexity of modern 3+ billion part designs Nvidia has changed tack with Kepler, releasing a simpler middle chip first, then scaling it both up and down. With the 600 series the Green team has debuted with an upsized 'middle chip' (GK-104) devloped it a smaller variant (GK-107) and are now evolving GK-110, the largest model.