Jen-Hsun Huang, Nvidia’s chief executive, disclosed during an annual conference that it will deliver a chip, code named Kepler, in late 2011 that will be three to four times the performance of Fermi with little increase in power consumption.
In 2013, Nvidia will make an even greater leap, Huang said, predicting the arrival of a chip called Maxwell will be ten to 12 times the power of Fermi.
Kepler will be made off the 28nm process, and will be focused far more heavily on performance per watt than Fermi is. Nvidia estimates that Kepler will be three to four times the performance per watt over Fermi, giving us a ballpark of five gigaflops of per watt. In 2013, the Maxwell chip should bring 16 times the performance in parallel graphics-based computing, as well as the ability to work independent of a CPU. Performance per watt will triple over the previous generation, hopefully hitting 15 gigaflops per watt. If Nvidia has the new chips top at a 300 watt thermal envelope, then the Kepler Teslas will deliver around 1.5 teraflops and the Maxwell Teslas will top out at around 4.5 teraflops.
The future Nvidia chips can go into the new SGI project mojo systems to deliver a petaflop in a single rack and then with Maxwell 2-3 petaflops in a single rack.
320 Maxwell chips would get about 1.44 petaflops.
The Tsubame is notable because it leverages GPU clusters, and its success is one of the milestones for GPU’s in supercomputing. The Tsubame 1.0 uses Tesla S1070 in a 680 GPU cluster. With it, scientists have been able to experience speed ups of up to 80x in problems like weather modeling. Coming in December, the next-generation Tsubame 2.0 will use the Tesla 2050 with 4224 GPUs and provide performance of more than 3 PFLOPs.