Proliferated growth in Machine learning counterparts challenges Silicon Technology.
The rise of artificial intelligence and impending end of Moore’s law means silicon chips are nearing the end of the line. Here are some alternatives.
SILICON has been making our computers work for almost half a century. Whether designed for graphics or number crunching, all information processing is done using a million-strong horde of tiny logic gates made from element number 14.
But siliconâ€™s time may soon be up. Mooreâ€™s law â€“ the prophecy which dictates that the number of silicon transistors on microprocessors doubles every two years â€“ is grinding to a halt because there is a limit to how many can be squeezed on a chip.
The machine-learning boom is another problem. The amount of energy silicon-based computers use is set to soar as they crunch more of the massive data sets that algorithms in this field require. The Semiconductor Industry Association estimates that, on current trends, computingâ€™s energy demands will outstrip the worldâ€™s total energy supply by 2040.
So research groups all over the world are building alternative systems that can handle large amounts of data without using silicon. All of them strive to be smaller and more power efficient than existing chips.
Julie Grollier leads a group at the UMPhy lab near Paris that looks at how nanodevices can be engineered to work more like the human brain. Her team uses tiny magnetic particles for computation, specifically pattern recognition.
When magnetic particles are really small they become unstable and their magnetic fields start to oscillate wildly. By applying a current, the team has harnessed these oscillations to do basic computations. Scaled up, Grollier believes the technology could recognize patterns far faster than existing techniques.
It would also be less power-hungry. The magnetic auto-oscillators Grollier works with could use 100 times less power than their silicon counterparts. They can be 10,000 times smaller too.
Igor Carron, who launched Paris-based start-up LightOn in December, has another alternative to silicon chips: light.
Carron wonâ€™t say too much about how his planned LightOn computers will work, but they will have an optical system that processes bulky and unwieldy data sets so machine learning algorithms can deal with them more easily. It does this using a mathematical technique called random projection. This method has been known about since 1984, but has always involved too many computations for silicon chips to handle. Now, Carron and his colleagues are working on a way to do the whole operation with light.
â€œOn current trends, computingâ€™s energy demands could outstrip total supply by 2040â€œ
What will these new ways of processing and learning from data make possible? Carron thinks machines that can learn without needing bulky processors will allow wearable computing to take off. They could also make the emerging â€œinternet of thingsâ€ â€“ where computers are built into ordinary objects â€“ far more powerful. These objects would no longer need to funnel data back and forth to data centres for processing. Instead, they will be able to do it on the spot.
Devices such as Grollierâ€™s and Carronâ€™s arenâ€™t the only ones taking an alternative approach to computation. A group at Stanford University in California has built a chip containing 178 transistors out of carbon nanotubes, whose electrical properties make them more efficient switches than silicon transistors. And earlier this year, researchers at Ben-Gurion University in Israel and the Georgia Institute of Technology used DNA to build the worldâ€™s smallest diode, an electronic component used in computers.
For the time being, high-power silicon computers that handle massive amounts of data are still making huge gains in machine learning. But that exponential growth cannot continue forever. To really tap into and learn from all the worldâ€™s data, we will need learning machines in every pocket. Companies such as Facebook and Google are barely scratching the surface. â€œThereâ€™s a huge haul of data banging on their door without them being able to make sense of it,â€ says Carron.
Editor’s note: Original Source:Â ‘NewScientist’
This article appeared in print under the headline â€œMaking light work of AIâ€
Hal Hodson. “Move over silicon: Machine learning boom means we need new chips”
NewScientist. N.p., Web. 24Â August. 2016.