A Look at Intel's 80-Core Processor Hitting Two Teraflops

I got to see the 80-core Polaris research processor in action at today's Research at Intel Day presentation, where Intel shows off some of its latest and greatest tech.

As I watched Polaris crank out a whopping 2 teraflops (that’s two trillion floating point calculations per second), I have to say I semi-consciously waited for the room to vibrate, or blaze with heat, or maybe leap back in time. But no, the little sucker just quietly did its work. Granted, it was a relatively noisy demo room, but the only real indication that it was pumping out this insane processing power was the readout on the screen showing each of the 80 cores as green.



(Click for a larger image)

In addition to crunching the actual numbers, Polaris moves them around between cores at the same breakneck pace. According to Nitin Borkar (pictured below), the chip has terabytes of core-to-core bandwidth to match its teraflops of processing.

Borkar first pushed the chip to 1 teraflop, which ran each core at 3.13Ghz. To hit 2 teraflops, each core bumped up to just over 6Ghz - but that also used about 4 times as much power.

Each chip measures 66mm by 66mm, Borkar says, though I couldn't see the individual chip underneath its heat sink. He did show an example wafer from which the chips are cut, though. Each of the somewhat-visible individual rectangles in this dinner-plate sized wafer packs in 80 cores, along with memory that is laid on top of the cores in the chip itself.



(Click for a larger image)

Polaris is impressive to say the least. But it's something of an idiot savant in that it's really, really good at one type of thing - in this case, pushing out floating point calculations - but it can't perform the full range of functions you'd get with a regular CPU. You can't run Windows on it, for instance. That's why it's billed as a research chip.

Comments

Post a comment Post a comment

0 / 1000 max characters