OPTICAL CHIPS WERE DEVELOPED AND TESTED SEVERAL YEARS AGO, BUT WITH THE RISE OF DEEP LEARNING, THEIR FUNCTIONALITY CAN CHANGE THE STATE OF THINGS.
Inside a tiny Boston Harbor District lab, buried in a jumble of lasers, lenses, mirrors, and a tangle of cables, is a tiny chip that could have a big impact on the world of artificial intelligence.
The laboratory belongs to Lightelligence, a startup developing a radically new type of AI compute acceleration chip. Instead of using electrons to do the basic mathematical calculations needed for machine learning, the company’s prototype device uses light.
In theory, the transfer of information using the properties of light is much faster than electric current and could allow AI algorithms to run hundreds of times faster than the best AI microprocessors known today. Since computing power, especially in learning deep neural networks, is significant in terms of competitive advantage, this could mean a huge leap forward in the development of complex algorithms and systems. In all cases, the speed of computation is also linked to the interfacing with other peripherals, memories, buses, chipsets and all the systems operating in a conventional manner. Lightelligence must therefore also create guidelines to take full advantage of its microprocessor, improving its characteristics.
The deep neural network is spreading across all industries precisely because of its ability to constantly improve even in the absence of human control. In the early stages of learning, however, a lot of interaction is needed, just to carry the neural network, and a lot of computing power.
THE GROWTH OF DEEP LEARNING HAS ALREADY SPARKED A FLAME OF COMMERCIAL ACTIVITIES AROUND NEW CHIPS OPTIMIZED FOR COMPLEX MATHEMATICS.
Twenty-year-old CEO Yichen Shen of Lightelligence illustrated how light and its use in the AI industry have huge competitive advantages. Photons are faster than electrons and their movement through the circuits of a microprocessor, without jaule effect, eliminates the problem of overheating and relative loss of energy, with considerable incremental capacity in the computation, also reducing the spaces of the data center or the supercomputer.
It should be noted that the calculations using light are very demanding, in fact the optical chips developed previously have not given sufficient results precisely because of the difficulty of predicting the behavior of photons in emulation of transistors.
IN THE AGE OF DEEP LEARNING, LINEAR CALCULATIONS IN WHICH OPTICAL DEVICES EXCELLENT ARE SUITABLE FOR EXECUTING MATERIAL MULTIPLICATIONS NECESSARY FOR DEEP LEARNING.
Shen and his Lightelligence collaborators are currently working with a company that produces microprocessors on a large scale precisely to test this innovative technology, projected into the future for commercial evolution.
Other actors like UCLA they sift the same ground, the light, exploiting in this case the refraction of polymers made with 3D printers. Details of the device, called a diffractive deep neural network (D2NN), have been published in the journal Science, and UCLA professor Aydogan Ozcan explains that some of these experiences lead to the design of components and systems that work differently from traditional systems, and that may be more convenient to market than in the past.
Dan Hutchinson, an analyst at VLSI Research who tracks innovative microprocessor designs, says interest in new optical chips is growing with advancements in the design and manufacture of devices used for the network. Optical chips are also relatively easy and inexpensive to manufacture., and can reduce the barrier to entry for startups.
However, Lightelligence still faces major challenges. Zhangxi Tan, a chip industry veteran and CEO of another microprocessor start-up, OURS Technology, doubts the new process could prove difficult to integrate into large-scale systems.
In particular, the difficulty lies in the fact that there are no tools suitable for the design of optical microprocessors, in fact both software and for example lase drivers, electronic modulators and other necessary components, do not are not very suitable for this type of research. The challenge will therefore not only be in the development of the unique element, but of a whole series of structures necessary to improve the design and development processes.
HOWEVER, THE EFFORT IS QUICKLY MOVING UPON CHANGES.
Just last year, Shen was a doctoral student studying photonic materials in Marin Soljacic’s lab at MIT. With Soljacic and several other students, he published an article in the journal Nature Photonics describing a new way to perform neural network calculations using optical interference. The idea of starting a business, however, predates the publication of official documentation, thanks to a West Coast venture capitalist.
Lightelligence has a direct competitor in this area, Lightmatter. The CEO of Lightmatter is not new, on the contrary he participated in the project with Shen and he too managed to obtain similar funds for the development of optical chips in deep learning. All the players are convinced that a healthy rivalry could help accelerate the development of this technology.