Switching sees the light
Photonic interconnects will accelerate data transmission by enormous magnitudes at botdevice and component levels.
The advantages of using photonic technology rather than electrical circuits for data transmission within computer systems have long been apparent to the IT industry. But, to date at least, the cost of manufacturing the tiny optical components required has presented an insurmountable barrier to the technology's integration.
That looks set to change over the next few years though: increasing demand for faster input/output (I/O) bandwidth and scalability, particularly in server architecture, has persuaded leading IT companies and academic organisations to put more time and money into developing hybrid silicon photonic components.
Stan Williams is a senior fellow and founding director of the Quantum Science Research Group at HP Labs, which is investing millions of dollars in researching lower-cost photonic interconnects for its servers, supercomputers, and data centre products.
"We have known for 300 years that photonics are better than electronics at moving information around, as exampled by optical communications in telephones," Williams says. "The problem for computers has been that they are too expensive."
He estimates that at present, a single rack of photonic interconnects, which might contain up to 40,000 lasers, would cost hundreds of thousands of dollars, making optical communications the major part of any system cost.
"Only one or two customers might be willing to pay that for the increased performance; but for most, it is out of sight," Williams admits. "We need to bring that cost down by a large amount."
For the moment, the components within servers and high-performance computers are linked by copper wires embedded either within motherboards or cables. Rack and blade servers, for example, are crowded with copper lines that drain power, generate heat, and force system designers to put hot components right next to each other, further exacerbating cooling problems. Because heat and cooling problems limit the number of processors that can be used within the system, performance is rarely maximised, while the electromagnetic interference caused by sending electrical pulses across system boards can also cause problems.
Most experts agree that, as CPU-intensive applications drive demand for tera-scale computing using larger numbers of multi-core processors, those copper wires will soon become overloaded.
"The interconnect challenge comes from how many wires need to send messages among the system components," says Williams. "As the number of messages gets larger [due to CPU intensive applications], the numbers of wires gets very large, while memory and I/O bandwidth are limited by the number of pins. The bit-rates are limited by the device speed, and the pin counts are not increasing."
According to HP, Intel, and others, the best way to deliver the I/O bandwidth and system speed to overcome these limitations is to use optical technology, but combine it with the cheaper component manufacturing processes associated with traditional complementary metal oxide semiconductor (CMOS) techniques: hybrid silicon lasers, or silicon photonics.
These transmit data optically by sending light rather than electrical pulses through silicon, or at least through a layer of material bonded to the silicon, which can actually catch and process the light. HP is exploiting its expertise in nanoimprint lithography to provide this layer in combination with a silicon ring resonator; other players are exploring other materials like indium phosphide to achieve the same end.
"If we put the optical transceivers on the region of the motherboard that really needs the data, and beam the data directly between the two boards, we can really change the way in which computing can be done," Williams believes. "The decrease in latency alone should be able to speed things up in an order of magnitude because photons are much faster, and use much less energy."
HP has developed an optical data bus that uses a single milliwatt of laser power to connect eight different modules at 10Gbps per channel, giving an aggregate bandwidth of over 250Gbps. IBM has performed similar research in collaboration with hybrid silicon optical specialist IntexyS Photonics. In 2006, the partnership produced opto-electronic modules capable of 120Gbps of aggregate system bandwidth. The conversion from electrical signals to optical signals was carried out by an array of 12 lasers, each of which could process 10Gbps of data per second.
In December 2008, Intel's Photonics Technology Lab demonstrated a silicon photonic integrated chip (PIC) capable of transmitting data at aggregate speeds of 200Gbps. This was done using an external multi-wavelength laser source, however, which will need to be replaced with a hybrid silicon laser array if higher terabits per second bandwidth can be possible. Intel developed a silicon modulator that can encode data on an optical beam at a rate of 40Gbps in 2007.
Many vendors, including HP, as well as academic research organisations, are working on developing the laser arrays needed to transmit those photons. And, in most cases, they are focusing on adapting LED transmitters to do it at less cost.
"We need to decrease the cost of light sources within the system, and we have laser and LED engineers trying to simplify component construction while maintaining performance," says Williams. "We do not need the fancy laser packages that people use for telecoms - we can tweak up LED performance so they might be able to do the job."
Similar development is going on at the UK Institute of Photonics (IOP), a commercially-orientated research unit within the University of Strathclyde. Technologies under analysis include III-V compound semiconductors, and LEDs for use in optical data storage and telecommunications.
"We are taking semiconductor wafers, and turning them into arrays that contain thousands of entities on one rack for board-to-board communications, supporting nanosecond switching times on each LED," says IOP business development manager, Simon Andrews. "If you have thousands of those running in parallel it gives you a fantastic data-rates in a variety of wavelengths."
Some experts argue that using photonics as the basis for short-range LANs within the data centre may be stretching the technology's potential capabilities too far, but HP remains adamant that free space server-to-server communications are possible. Explains IOP's Williams: "There are free space optical interconnects, at all length scales from millimetres to multiple metres for blade-to-blade connections. We can rip the copper out of the data centre and replace it with nothing but light travelling through air."
HP initially envisages moving data optically along the motherboard both onto and off the available CPUs, so that, while the calculations are performed within the silicon processor core or some level of cache memory, everything else is done with light. The communications bandwidth can easily be a thousand times higher, because there are no transistors waiting for data to arrive before they can operate at full load.
Ultra-bright crystal GaN photonic LEDs
Nobody is saying that computers will ever operate solely on photonic communications, or that it will ever be cost-effective to put optical interconnects into portable devices (arguably where they are most needed given the lower power requirements and ongoing drive to component miniaturisation); but many other potential uses within mainstream ICT and computing applications are possible, according to Andrews.
For example, a UK partnership between Sharp Laboratories Europe, the University of Glasgow, the IOP and Mesophotonics has just kicked off a $2.4m project to build ultra-bright crystal GaN photonic LEDs for LCD backlights. It is thought that silicon photonics may also be applied to optical debug of high speed data and expanding wireless networks by transporting analogue RF signals.
There is also strong interest in semi-conductor disk lasers being tested by the IOP, essentially a different way of producing an array of lasers that could again support new forms of parallel communications technology, says Andrews: "These can bring down the cost of those components by using different materials and grow laser gain media to give exactly the wavelength you want. These are not expensive crystals but one wafer that can be diced into thousands".
And, as HP and Intel are focusing primarily on silicon photonics technology into server backplanes and chip to chip data centre communications in the short term, they and other manufacturers including IBM are looking to develop on-chip solutions where the processor itself sends data around the chip using optical communications.
With what seems to be so much in the way of research resources being put into silicon photonics, it seems inevitable that components based on the technology will soon find their way into commercially available systems. IOP's Williams believes that photonics products could debut as soon as this year and appear in volume within the next five years. First, because the silicon manufacturing base is already in place; and second, because HP's photonic technology will be designed to integrate with existing silicon-based systems, which are already widely understood within the IT industry. "Everything we have done is CMOS compatible, and with the exception of the lasers and LEDs themselves, it is all silicon-based," he adds. "We have not thrown away any of the existing infrastructure, just added more features onto the chip and it is our intention from an architectural level to make it as transparent as possible."
Any delay in the appearance of commercial products may have less to do with ongoing technology or implementation challenges as opposed to big vendor politics and the will to avoid cannibalising their own revenues from silicon based products.
"The properties of different optical wavelengths, as well as miniaturisation, are now being explored, and are catching up [to silicon] rapidly," the IOP's Williams reports, "but there are still plenty of other factors coming into play here."
Industry insiders have been impressed by the fact that the larger IT companies do appear to be pulling in the same direction at least; but whether the presence of a common goal leads to the emergence of standardised, compatible silicon photonic components remains to be seen.