Tech
Intel looks to fibre-optics to speed chips
The chipmaker is checking out a new approach to building processors that uses fibre-optics to connect components, and even eventually to replace copper or aluminium wires inside chips
When it comes to developing faster computers, Intel has seen the light.
The chipmaker recently launched a new effort inside its Microprocessor Research Lab aimed at examining a move from using aluminium and copper wires for moving data inside computers to using fibre-optic strands. Fibre could be used to connect components, researchers say, but it could also eventually replace wires inside chips.
Power and speed are the main motives. Designers are faced with the conflicting goals of shrinking chips and simultaneously adding more transistors to boost performance. Adding transistors means increasing the electricity required to run a given chip. Unfortunately, shooting more power through a smaller chip increases signal interference, cuts down on battery life in notebooks and leads to more manufacturing defects.
Optical connections, which use laser light beams rather than electrical impulses to transmit signals, could solve some bottlenecks because they can run on less power. It won't be cheap or easy, but it might become the lead alternative. Other companies are also looking at organic connections.
"There are barriers for how much you can lower the voltage," said Bill Pohlman, chief executive of Primarion, a chip designer working on ways to integrate optical fibres into semiconductors. "We're talking about a billion transistors on a chip operating at a half a volt at well over 10GHz."
In addition to requiring less power, Fibre is also faster. Copper wire, the current medium for connecting PC components, offers data transfer rates of about 10gbit/s (gigabits per second), and Intel expects rates to reach 20gbit/s over the next few years.
The speed increase will buy time for copper, but optical interconnects will run at 30gbit/s to 60gbit/s.
Chip design hitting a wall
The optical push comes as part of an industry-wide concern about the future of processor design. For years, manufacturers, obsessed with performance, added transistors and boosted clock speed. Now however, the industry is butting up against the laws of physics. Transistor bloat is creating a situation where extreme amounts of power are being corralled into tiny spaces, and this puts pressure on the interconnects. "Although it's much less publicized, as megahertz rises, signalling (the speed at which interconnects can share data) has to keep track with it," said Wilfred Pinfold, technology director for Intel's MRL. "It's been an ongoing and important focus in the lab to keep signalling moving forward." Intel researchers, as well as analysts and other industry sources, believe optical interconnects could begin to replace copper for data transmission between computers and computer components over the next five to seven years. Initially, fibre research will go into looking at how to link PC components over short distances. Then it'll delve into smaller components. "We will be pushed to develop optical interconnects," said Pohlman, who also used to be an Intel executive. "Optical is attractive to me because bandwidth can be dramatically scaled to (follow) Moore's Law," Pohlman said. "Buses of 30gbit/s per sec and 60gbit/s will be possible." Moore's Law states that the number of transistors on a microprocessor doubles approximately every 18 months. The 10GHz fibre-optic chip
Connecting components is only one application for fibre. Over the longer term, possibly 10 years, optical interconnects could also work their way into processors themselves, where copper wiring is just starting to become standard. Envisioneering Group analyst Rick Dougherty says that because of their bandwidth capabilities and lower-power characteristics, optical connections could be used within semiconductors to help push chips to higher clock speeds, approaching 10GHz. "Definitely, the power handling and optical interconnects are the two biggest issues" facing chip designers, he said. Additionally, in laying down fibre, the industry could leverage its years of experience in making chips out of aluminium wires. "Its easier to do optical structures in aluminum than in copper," Dougherty said. And both Intel and rival AMD have a lot of expertise in this area. But Pinfold called optical's move into internal chip interconnects, which link transistors, "controversial" at the moment. That's because it's cheaper to run a number of copper wires than a single optical fibre, because of the cost and complexity of adding a laser on one end and a receiver on the other to transmit optical signals. But optical connections could be used in areas such as clocking to help eliminate the electrical clocking delays present in current chips. It could also be used where circuits need complete electrical isolation. "We have not yet come to a position where we need to use either of those, but we're certainly looking at them for the future," Pinfold says. "Somebody has to show me what the compelling reason is to do it. On-chip (optical) is never going to be substantially cheaper than with copper." Price is the limiting factor
Pohlman agreed that cost is a major limiting factor. "The holy grail is low-cost electrical-to-optical conversions," Pohlman said. Optical interconnects are bound to cost more than copper wiring because of the cost of manufacturing the lasers and receivers needed to translate data from an electrical signal, send data over the fibre and translate it back. But costs have come down over time, pushing fibre from long-distance communications to metropolitan networks. More recently, companies have begun using optical for local area networks, the small-scale networks used internally by businesses. The Cahners In-Stat Group reports that optical networking is growing at a compound annual rate of 42 percent. Much more work needs to be done, however, to make optical connections cheap enough for widespread use inside computers. "A lot of research is going into lower-cost lasers, lower cost-receivers," Pinfold said. Miniaturising these components and manufacturing with readily available processes will be a major undertaking for Intel and other companies. One way to reduce cost will be to create a way of manufacturing lasers and receiving equipment using CMOS, Intel's standard silicon chipmaking technology. Several universities, including the Massachusetts Institute of Technology, Stanford University and the University of Surrey, are working on creating CMOS lasers. "It's still a long way off because of its cost issue," Pinfold said. For that reason, the "crossover is still fuzzy." Breakthroughs in copper could push optical out even further, he said. However, optical connections are almost inevitable, said Tony Massimini, analyst with Semico Research. "The question is not 'Will it happen?' It's 'When?'" Massimini said. "We're talking about mainstream, high-volume production. And it's going to take a number of years to get to that point. It's one thing to have something working in the lab. It's another thing to do it in mass production." Though he says it will likely take five to 10 years for optical connections to work their way from networking separate computers down to connecting elements inside a chip, Massimini sees the optical transition as an inevitable one. "We'll see that the first uses of it will be chip-to-chip on a board. Once people start working with that...the next step is to use it internally" inside a chip, he said. It's always a question of when. But certainly something like this, when it starts, it's like a snowball...It starts slowly, then hits a critical mass." CNET News.com's Michael Kanellos contributed to this report. See Chips Central for the latest headlines on processors and semiconductors. See the Hardware News Section for full coverage. Have your say instantly, and see what others have said. Click on the TalkBack button and go to the Chips Central forum Let the editors know what you think in the Mailroom. And read other letters.
The optical push comes as part of an industry-wide concern about the future of processor design. For years, manufacturers, obsessed with performance, added transistors and boosted clock speed. Now however, the industry is butting up against the laws of physics. Transistor bloat is creating a situation where extreme amounts of power are being corralled into tiny spaces, and this puts pressure on the interconnects. "Although it's much less publicized, as megahertz rises, signalling (the speed at which interconnects can share data) has to keep track with it," said Wilfred Pinfold, technology director for Intel's MRL. "It's been an ongoing and important focus in the lab to keep signalling moving forward." Intel researchers, as well as analysts and other industry sources, believe optical interconnects could begin to replace copper for data transmission between computers and computer components over the next five to seven years. Initially, fibre research will go into looking at how to link PC components over short distances. Then it'll delve into smaller components. "We will be pushed to develop optical interconnects," said Pohlman, who also used to be an Intel executive. "Optical is attractive to me because bandwidth can be dramatically scaled to (follow) Moore's Law," Pohlman said. "Buses of 30gbit/s per sec and 60gbit/s will be possible." Moore's Law states that the number of transistors on a microprocessor doubles approximately every 18 months. The 10GHz fibre-optic chip
Connecting components is only one application for fibre. Over the longer term, possibly 10 years, optical interconnects could also work their way into processors themselves, where copper wiring is just starting to become standard. Envisioneering Group analyst Rick Dougherty says that because of their bandwidth capabilities and lower-power characteristics, optical connections could be used within semiconductors to help push chips to higher clock speeds, approaching 10GHz. "Definitely, the power handling and optical interconnects are the two biggest issues" facing chip designers, he said. Additionally, in laying down fibre, the industry could leverage its years of experience in making chips out of aluminium wires. "Its easier to do optical structures in aluminum than in copper," Dougherty said. And both Intel and rival AMD have a lot of expertise in this area. But Pinfold called optical's move into internal chip interconnects, which link transistors, "controversial" at the moment. That's because it's cheaper to run a number of copper wires than a single optical fibre, because of the cost and complexity of adding a laser on one end and a receiver on the other to transmit optical signals. But optical connections could be used in areas such as clocking to help eliminate the electrical clocking delays present in current chips. It could also be used where circuits need complete electrical isolation. "We have not yet come to a position where we need to use either of those, but we're certainly looking at them for the future," Pinfold says. "Somebody has to show me what the compelling reason is to do it. On-chip (optical) is never going to be substantially cheaper than with copper." Price is the limiting factor
Pohlman agreed that cost is a major limiting factor. "The holy grail is low-cost electrical-to-optical conversions," Pohlman said. Optical interconnects are bound to cost more than copper wiring because of the cost of manufacturing the lasers and receivers needed to translate data from an electrical signal, send data over the fibre and translate it back. But costs have come down over time, pushing fibre from long-distance communications to metropolitan networks. More recently, companies have begun using optical for local area networks, the small-scale networks used internally by businesses. The Cahners In-Stat Group reports that optical networking is growing at a compound annual rate of 42 percent. Much more work needs to be done, however, to make optical connections cheap enough for widespread use inside computers. "A lot of research is going into lower-cost lasers, lower cost-receivers," Pinfold said. Miniaturising these components and manufacturing with readily available processes will be a major undertaking for Intel and other companies. One way to reduce cost will be to create a way of manufacturing lasers and receiving equipment using CMOS, Intel's standard silicon chipmaking technology. Several universities, including the Massachusetts Institute of Technology, Stanford University and the University of Surrey, are working on creating CMOS lasers. "It's still a long way off because of its cost issue," Pinfold said. For that reason, the "crossover is still fuzzy." Breakthroughs in copper could push optical out even further, he said. However, optical connections are almost inevitable, said Tony Massimini, analyst with Semico Research. "The question is not 'Will it happen?' It's 'When?'" Massimini said. "We're talking about mainstream, high-volume production. And it's going to take a number of years to get to that point. It's one thing to have something working in the lab. It's another thing to do it in mass production." Though he says it will likely take five to 10 years for optical connections to work their way from networking separate computers down to connecting elements inside a chip, Massimini sees the optical transition as an inevitable one. "We'll see that the first uses of it will be chip-to-chip on a board. Once people start working with that...the next step is to use it internally" inside a chip, he said. It's always a question of when. But certainly something like this, when it starts, it's like a snowball...It starts slowly, then hits a critical mass." CNET News.com's Michael Kanellos contributed to this report. See Chips Central for the latest headlines on processors and semiconductors. See the Hardware News Section for full coverage. Have your say instantly, and see what others have said. Click on the TalkBack button and go to the Chips Central forum Let the editors know what you think in the Mailroom. And read other letters.