IBM Compresses 100Gbps Network Onto A CMOS Chip

IBM researchers have made cheap low-powered analogue to digital converters (ADCs) which could allow 100Gbps networks over long-distance fibre, with a cheap device to send and receive data each end.

The breakthrough will allow cheaper and simpler devices that can convert the signals on fibres into digital information. This will cut energy use and could revolutionise mobile phones and radio astronomy, while making high speeds easier and cheaper to deliver to a wider range of devices. The technology, produced by IBM researchers, working with Ecole Polytechnique Fédérale de Lausanne (EPFL) in Switzerland, is based around tiny ADCs integrated onto the same CMOS chips which hold other functions.

Smaller ADCs than anyone else

Although light signals are sent along fibres in digital form, by the time they reach the far end of the cable they have to be processed to clean up and receive the data which was sent, explained Martin Schmatz, who manages the systems division of IBM’s Zurich lab: “The problem is fibre has dispersion,” he explained, so the signal is blurred when it reaches its destination. “During transit, some energy is dispersed to other frequencies. We need to shift that bit of energy back to the right place.”

Currently, long distance links use “dispersion compensating” fibres, in which the main fibre is combined with another which has the opposite dispersion characteristics. This is complex and expensive says Schmatz – and the IBM breakthrough allows the signal to be cleaned up with dispersion cancelled out electronically.

“You can actually correct for dispersion by a mathematical approach,” Schmatz told TechWeekEurope. “If you digitise the signal, you can then apply mathematical functions to correct for the dispersion.” Telephone lines regularly use maths to correct for signal dispersion but – as Schmatz pointed out – they only have to work at around 56kbps on digitised voice traffic. Digital correction on a fast fibre network would have to work far faster.

Fibre networks carry long-haul Ethernet data at 100Gbps, but to view them as analogue signals and clean them up, needs an eight bit resolution, said Schmatz: “If you have four channels, all of a sudden you have 2.5 Terabits persecond (Tbps) coming out of the ADC.”

ADCs have normally been implemented as separate components made with different technology, because traditional CMOS circuits aren’t optimised for analogue signals. However,  it is expensive and inefficient to take data at that rate off one chip and onto another, Schmatz told us.

The IBM team managed to show it is possible to make a very tiny ADC out of standard 32nm CMOS, opening the way to integrating many of these components onto the same chip as the rest of the network hardware.

“It is power inefficient to ship the data from an ADC to a CMOS chip,” he said. “We were able to show that it is possible to use a plain vanilla digital process on a digital CMOS chip, not optimised for analogue, to build such an ADC which has an extremely high content of analogue circuits.”

These ADCs are very power efficient, and consume a tiny amount of the CMOS chip’s area: “The majority of the power – and area – needs to be assigned to the digital circuit. The ADCs need to be tiny.”

100Gbps Ethernet signals are sent as four 25Gbps streams, separated by phase and polarisation.  Because the Nyquist-Shannon theorem says they need “oversampling”, a 100Gbps Ethernet channel would actually need four 64Gbps ADCs.  That’s faster than can be easily done, so the group proposes using more ADCs, each of which handles a time slice of the total signal (so it might use 64 ADCs, each at 1Gbps).

“We will end up with 256 ADCs on one side of the chip and a lot of processing on the other,” predicted Schmatz. “Is this possible? Absolutely!”

Radio telescopes and phones

The technique could be available in 2014, he said, and will have applications beyond fast long-haul Ethernet.”There are many applications where the signal you are using is in the analogue domain,” he said.

For instance, a set-top box could take in the whole frequency spectrum from a cable modem and sort out the individual channels in a single chip.

The approach could also revolutionise radio astronomy, which is all about finding signals in a big analogue spectrum. Schmatz hopes to provide signal processing equipment for the Square Kilometer Array (SKA), an international project to build the world’s largest and most sensitive radio telescope.

Another possibility would be to have one chip to handle all the radio signals coming into a mobile phone. “You could have one wideband antenna, sample everything from 400MHz to 6GHz – and then sort it out by number crunching  so you have Wi-Fi and Bluetooth at 2.4 GHS, and 3G and 4G at frequencies like 900MHz – and you do all of that in the digital domain.”

“Most of the ADCs on the market today weren’t designed to handle the massive Big Data applications we are dealing with today – it’s the equivalent of funnelling water through a straw from a fire hose.”

How much do you know about IBM? Try our quiz to find out!