Codenamed Corona, this laser-powered contraption would handle 10 trillion floating points operations a second. In other words, if you put just five of them together, you’d approach the speed of today’s supercomputers. The chip’s 256 cores would communicate with each other at an astonishing 20 terabytes per second, and they’d talk to memory at 10 terabytes a second. That means it would run memory-intensive applications about two to six times faster than an equivalent chip made with good old fashioned electric wires.
More importantly, Corona would use a lot less power, helping the world’s supercomputers break the vaunted exascale barrier — i.e. deliver a machine that cab handle one quintillion (10 to the 18th) floating point operations a second. That’s 100 times faster than today’s fastest supercomputer. “Electronics…cannot scale to the scale that we need for these large systems,” says HP Labs researcher Marco Fiorentino.
This sort of optical chip communication is known as “integrated photonics”. Telecommunications networks and high-speed computer interconnects already use light to send information faster and more efficiently — think “fibre optics” — and now, HP and other research outfits are pushing to use light for communicating between computer computer chips or even between components built into the chips themselves.
Corona is just one of several efforts to build superfast chips that can bust through the exascale barrier, including Intel’s Runnemede, MIT’s Angstrom, NVIDIA’s Echelon and Sandia’s X-calibur projects. All seek to use integrated photonics in some way, but the technology is the heart of the matter for HP’s 256-core Corona.
The catch is some of the technology needed to build Corona doesn’t exist. But that’s changing. Recently, researchers and chip makers have shrunk optical communications devices so that they can be put onto chips. They’ve made chip-scale equivalents of cables, modulators and detectors. “A lot of people have concentrated on individual devices,” said HP’s Fiorentino. “Now they’re starting to build circuits. It’s like going from the transistor to the integrated circuit.”
Fight the Power
There are two roadblocks preventing us from continuing to scale up the performance of today’s chips at the current rate. The more processor cores we cram on each chip, the more challenging it is to coordinate them. And as computer systems get bigger, moving data in and out of memory becomes a huge energy drain. Integrated photonics can help with both problems by providing high speed, low-power communications.
When you get beyond 16 cores per chip, it becomes very difficult for the chip to function as a parallel processor without the cores being able to communicate with each other, says Lionel Kimerling, a materials science and engineering professor at MIT. “There’s going to be no way to scale performance without some sort of broadcast or near-broadcast capability,” he says.
The aim is to build a tiny laser into each core, so that it can broadcast information to all other cores through an optical network. With even a minimal level of communication among processors, you can ensure uniform heat dissipation across the chip, and you can ramp clock speed up and down depending on workloads. This will not allow us to reach unprecedented speeds, it will significantly reduce power consumption.
Using electronics for a 10 terabytes per second channel between a CPU and external memory would require 160 watts of power. But HP Labs researchers calculate that using integrated photonics lowers that to 6.4 watts.
Energy efficiency is a major issue for today’s servers, especially in large data centres that deploy thousands at a time. Right now, the major obsolescence factor for servers is power usage. The money saved on energy justifies buying a new server about every three years, says Kimerling. But integrated photonics, he says, could change that.
Integrated photonics is also likely to play a central role in boosting the bandwidth and lowering the power consumption of the internet, particularly for supporting video services. Mobile devices are also power constrained. And electromagnetic interference — something you don’t get with photonics — is a growing concern for mobile devices and automobile electronics. All of these technologies are eventually going to require integrated photonics, says Daniel Blumenthal, an electrical and computer engineering professor at the University of California, Santa Barbara. “Business just can’t be done the same old way.”
The Missing Piece
The missing piece of the puzzle is a way to generate light: the on-chip laser. Semiconductor lasers have been around for years and are widely used in telecommunications gear, laser printers and DVD players. These lasers are similar to computer chips and they’re small, but not nearly small enough to be used as light sources for optical circuits built into computer chips. For that, you need to make microscopic lasers as part of the chip-making process.
You can’t make a laser from silicon, so researchers around the world have been making lasers from other semiconductor materials that are more or less compatible with standard chip-making processes. These are usually indium phosphide or gallium arsenide. This is the approach Intel, HP and UC Santa Barbara are taking.
MIT’s Kimerling recently came up with a new approach: germanium. The material produces a laser that emits light at the wavelength used by communications networks, it operates at up to 120C and germanium can be readily grown on silicon.
Kimerling coordinates an industry technology roadmap for integrated photonics at MIT. He says the timeframes companies are giving for when they need the technology have shrunk by about three years over the past year. “Many people said 2017,” says Kimerling. “Now it’s 2013, and we’ll take it today if you can give it to us.”
According to Kimerling, major semiconductor fab will be turning out integrated silicon photonic products last this year. The products are likely to be simple transceivers, but it shows that photonics is rapidly becoming a standard part of the chipmaking toolkit.
Photonics in 3D
The computer industry’s immediate need for integrated photonics involves getting data on and off chips, says Richard Otte, CEO of Silicon Valley chipmaker Promex Industries. Integrated photonics for connecting components on-chip is probably 10 years out, he says.
As these technologies evolve, researchers are also developing “through silicon vias” or TSVs. Otte calls TSVs “the dark horse in this data rate transmission race”. TSV are vertical connections that make it possible to stack chips. For example, memory chips can be stacked on top of processor chips.
There’s a great deal of interest in 3D devices because chips are generally very thin — on the order of 50 to 100 microns — and expanding vertically saves a lot of space. This is particularly important in mobile devices. It also shortens the length of the interconnects between the components, which saves energy. Stacking is a leading candidate for keeping Moore’s Law on track, and many designs for future high-performance chips are 3D. “If TSV technology develops rapidly, on-chip [photonics] will be delayed,” says Otte.
Corona actually combines the two idea. It’s a 3D chip that uses integrated photonics. Or at least, HP hopes it will be. Each chip is slated to have 256 general purpose cores organised in 64 four-core clusters, and the cores will be interconnected by an all-optical, high-bandwidth crossbar. The goal is to build the chip’s processor cores using a 16-nanometer chipmaking process. And that should be available 2017.
Images: HP Labs