Nov 18, 2010 10:00 GMT  ·  By

As days pass by, it becomes ever more clear that Nvidia is dead serious when it comes down to GPGPU computing, the company just entering into a DARPA founded competition to build a exascale computer chip against the likes of Intel, MIT and Sandia National Labs.

Dubbed Echelon, this chip was detailed by Nvidia's chief scientist, Bill Dally, who is heading the project, at the Supercomputing 2010 conference that is taking place right now in New Orleans, Louisiana.

Although, at this time, the chip is only a paper design backed up by some simulations, Nvidia says it will be able to process a floating point operation using only 10 picojoules of power, compared to the 200 picojoules required by Nvidia's current Fermi based professional graphics cards.

During the conference Dally said that eight of these cores will be packaged on a single streaming multiprocessor (SM), 128 SMs following to be used for building the Echelon chip.

As a result Nvidia would get an 1024 core graphics chip, each of these cores being able to handle four double precision floating-point operations per clock cycle compared to one double precision floating-point operation per cycle in today's Fermi cores.

This would mean that Echelon would reach 10 teraflops in performance, Echelon also using an advanced RAM memory architecture that can be dynamically configured in order to get data as close as possible to the processing elements (this 256 Mbytes memory can be broken up into as many as six levels of cache) reducing the need of moving data around the chip and thus reducing power consumption.

As far as programming such a chip goes, Dally says the future will bring us a change in programing models, Nvidia thinking this should be an evolution of CUDA, although OpenCL, OpenMP and Microsoft's DirectCompute are not excluded from the equation.

As I said before this new chip is developed to take part in DARPA's Ubiquitous High Performance Computing program, its goal being to build a prototype petaflop-class system into a 57 kilowatt rack prototype computer by 2014. (via EETimes)