Last year, Lightelligence, a photonic AI chip company backed by the M.I.T. team, announced an important development -- the successful development of the world's first Prototype photonic chip. That could mean a new era for the chip industry.
Optical computing is not a new concept. As a completely different from electronic computing technology, optical computing with photonic carrier for information processing, rather than rely on light hardware electronic hardware, replace the electric operation with light operation, good at fast parallel processing highly complex computational tasks, but it hasn't been to find the right application scenarios, and restricted by the traditional separate optical device used to control the light field a single huge disadvantage, optical design, optical computing always stay in the laboratory.
In recent years, electronic computing has become increasingly constrained by Moore's Law, and the improvement of storage density and computing speed of information technology carriers has become increasingly inadequate. As a result, some people have shifted their attention from "electricity" to "light" with faster speed and lower energy consumption.
Photonic chips are a hot field because AI requires so much data processing, and current silicon-based electronic chips are in sight for the future.
Back to this chip: For the two-and-a-half-year-old company, the prototype proves the pioneering idea some of the team made in 2017, published in the journal Nature Photonics, to use photons instead of electrons for AI computing. At that time, the entire photon-computing system they were developing in the lab occupied half the lab. On this prototype, the team successfully used the photonic chip to run Google Tensorflow's own convolutional neural network model to process the MNIST data set. This is a benchmark machine learning model that uses computer vision to recognize handwritten Numbers and is one of the most famous benchmark data sets in machine learning. In the test, more than 95% of the operation of the whole model is processed on the photonic chip. The test results show that the processing accuracy of the photonic chip is close to that of the electronic chip (over 97%). In addition, the time of completing matrix multiplication of the photonic chip is less than 1/100 of that of the most advanced electronic chip.
Unlike the logic used by general-purpose chips, deep learning systems spend most of their time on low-precision matrix multiplication, which is the most time-consuming and power-hungry of AI algorithms. Currently, AI accelerators are based on electronic computing, but combined with all the cutting-edge technologies of electronic computing, photonics chips can perform the two most important steps in AI computing: memory-to-cell data transfer and matrix computing itself. These two aspects of photonic chips have unique advantages. Thus, the concept of photonic AI chip came into being. Using photons to do matrix multiplication has a promising future in reducing power consumption and increasing speed, which is likely to bring the biggest opportunity in optical computing history. The Lightelligence photonic chip is designed to accelerate AI from a new perspective.
Although the photonic chip has been put into quantitative production on a small scale, it must be admitted that the whole new technology is still at a very early stage, and photonic AI chip companies are still facing great landing risks and technical challenges. However, in the pursuit of a higher level of machine intelligence, the demand for chip computing power will only increase. Photonic AI chips are expected to make up for the deficiency of electronic chips and surpass them in some special computing fields. The researchers integrated indium phosphide's (InP) luminescent properties and silicon's optical routing capabilities into a single hybrid chip. When a voltage is applied to the indium phosphide, the light enters the waveguide of the silicon wafer, producing a continuous laser beam that drives other silicon photonic devices. This silicon-based laser technology could make photonics more widely available in computers because the cost of making silicon on a large scale could be significantly reduced. Intel believes that while the technology is still a long way from commercialization, it believes that in the future dozens, if not hundreds, of hybrid silicon lasers could be integrated into a single silicon-based chip along with other silicon photonics components. This marks the start of low-cost mass production of highly integrated silicon photonic chips.
TRUNNANO (aka. Luoyang Tongrun Nano Technology Co. Ltd.) is a trusted global chemical material supplier & manufacturer with over 12 years experience in providing super high quality chemicals and Nano materials. The indium phosphide produced by our company has high purity, fine particle size and impurity content. Lower, please contact us if necessary.
