Nvidia has revealed that Maxwell, the successor to its Fermi GPGPU, will be combine with Nvidia's GPU for supercomputing applications, potentially leapfrogging the likes of AMD, Intel or IBM as early as 2013.
NVIDIA's Tegra General manager, Mike Rayfield, told Hexus that Maxwell will be the first end-product to use Project Denver, which has been confirmed as being more than just a simple GPU-on-Silicon project.
This will be, according to Rayfield, a far greater resource investment for the company than just licensing the design from ARM. Back in September, NVIDIA CEO Jen-Hsun Huang confirmed that Maxwell will offer a 16x improvement over Fermi (and not Tesla as some have reported).
In real terms, this means that Maxwell will offer 16GFLOPS per watt or a whopping 4TFLOPs in double precision in computing power (assuming that the each consumes up to 250W) and the jump in performance will be partly due to the inclusion of the Cortex A15 as part of the mix (as well as a move to 28nm manufacturing).
The implications are mind-boggling for the world of super computers; a PetaFLOP (one million GFLOPS) computer based on Nvidia's Technology would require a mere 250 Maxwell cores or just over three 42U racks (with 84 per rack) while an ExaFLOP would require around 3,100 such 42U racks with 260,000 cores.
Without accounting for the power overheads, an ExaFLOP computer would require 65MW of power and would be powerful enough to simulate the human brain from an electrochemical point of view.
Such a computer was expected to go live in 2018 but Nvidia's announcement could mean that, given the funds, an ExaFLOP monster could become a reality as early as 2014, a mere 18 years after ASCII RED became the first Teraflop Computer.
In comparison, the current leader in the Supercomputer Top500 list is the IBM Sequoia which is set to be launched in 2011 and will deliver 20 petaFLOPs with a power consumption of 6MW and 1.6 million cores.