News 

At last year's IDF in San Francisco, the company showed a prototype design that could theoretically achieve teraflop speeds - that's a trillion floating-point operations per second.

We were a little underwhelmed at the time since all that was shown was a non-working wafer which was merely being used for testing fabrication methods.

This year Intel's Justin Rattner and Pat Gelsinger gave a live demonstration of a working 80-core part, operating in excess of 6GHz and giving 2 teraflops of performance.

The 80-core part isn't an IA (Intel architecture) processor - it will never run Windows. It's a collection of far simpler specialised cores designed only for floating-point maths.

Nonetheless it was extremely impressive to see the test system - with its large water-cooled

ADVERTISEMENT

heatsink embedded in a custom-designed Foxconn motherboard - running a specialist test program solving partial differential equations and initially showing a speed of 1 teraflop at 3.13GHz.

Despite all the cooling, at this speed the processor was consuming only 46W of power - less than most standard desktop processors. Upping the speed first to just under 5GHz and finally to 6.2GHz, 2 teraflops performance was achieved at a power consumption of only 191W, about 1.5 times the consumption of current top-end Extreme-series Core 2 Quad CPUs.

Rattner went on to discuss future plans for many-core CPUs. The major problem is finding a way of attaching them to main memory without causing huge bottlenecks. Intel's solution is to go 3D and use a die-stacking system, placing the memory die directly on top of a multicore processing die, allowing each individual core to communicate directly with main memory. The fabrication technique, used in the 80-core processor involves a 'copper bumping' process, where small lumps of copper are used to form interconnects when the two dies are sandwiched together, transferring power to the upper wafer as well as data.

Rattner also speculated that future super-high performance, many-core processors used in general PCs would use Intel's silicon photonics system for extremely high bandwidth interconnection to other parts of the system.