If IBM has its way, QPI and HyperTransport are about to look about as fast as your mom’s 56K modem. It seems that copper interconnects are simply not good enough anymore, and IBM researchers have developed a way to incorporate nano-scale CMOS optical devices onto the same slab of silicon your old-and-busted electron based transistors are fabbed on. This allows chips to use pulses of light to communicate with each other rather than electrical signals to achieve some mind boggling bandwidth numbers.
The discovery is dubbed “Silicon Nanophotonics” and is part of IBM’s quest to build an exa-scale computer, that is to say a super cluster that can perform at least one million-trillion –or an Exaflop – calculations per second. For the uninitiated/ignorant among us, the pecking order is as follows: “kilo”, “mega”, “giga”, “tera”, “peta”, then finally “exa.” Perhaps some real world examples (unrealistically assuming perfect scaling, and ignoring the many controversies of FLOPS) would help put that further into perspective.
- 215,518 – Radeon 5970 (4.64 TFLOPS ea) = 1 ExaFLOP
- 15,873,016 – Phenom II x6 @ 3.2GHz (~63 GFLOPS ea) = 1 ExaFLOP
- 333,333,333 – Pentium 4 @ 3GHz (~3 GFLOPS ea) = 1 ExaFLOP
- 5,405,405,406 – Pentium 166MMX (~185 MFLOPS) = 1 ExaFLOP
Pointless large numbers aside, some of you might be thinking “Hey doesn’t Intel have that lightpeak thing happening as well?” The short answer to that is yes, but Lightpeak is intended for a whole other realm of interconnects. Rather than focusing on internal chip-to-chip communications, Lightpeak could perhaps be more appropriately viewed as a USB cable on crack. Big Blue has stated that they hope to use this technology outside the chip as well to connect modules and even racks of servers together (Goodbye Infiniband?).
All joking aside, this is some really cool technology that will be thrust more and more into the limelight in the coming decade. As silicon hits lithographic limits, this technology could serve as the ultimate “glue” for tying many small processor cores together, instead of taking the single (monster) die approach we see today. IBM claims that a single Nanophotonic transceiver channel occupies only 0.5mm^2 of silicon real estate, and an area of only 4mm^2 can already house enough transceiver equipment to throw data back and forth at over a terabit/second. That would mean that each link can handle almost 16 gigabits/second already.
For some perspective, a 3.2GHz QPI link transmits data at about 204.8 gigabits/second, and the latest and greatest HyperTransport 3.1 protocol @ 3.2 gHz can handle twice that.
Hit up the official press release link here for some more in depth details on this nifty light show.
IBM is sitting on some really sweet technology here that could be the breakthrough we’ve been waiting for to usher in a new era of high throughput computing. Sources close to the voices in my head say that we can expect a multi-core Cyrix Media GX4 chip w/ on-die VooDoo 6 graphics some time in 2018 using these multi-terabit optical interconnects, but at present I can neither confirm nor deny these rumors.S|A