Hard technology is the focus of technology competitions. Among them, semiconductor chips have attracted the attention of people across the country and even the world.

Few people know that in addition to "stuck" products such as SoC 5G chips and AI dedicated chips, there is also a semiconductor chip that also has great potential - spectrum chips.

In the physical world we live in, when light shines on matter, it will form a specific spectrum. Just like human fingerprints, the spectral characteristics of each substance are unique, and even the spectrum of a crop is different at different growth stages. Through the spectrometer, the spectral information of the object can be obtained for research and analysis.

The spectrum chip is actually a sensor composed of thousands of micro-spectrometers. With the help of the photoelectric effect of semiconductor materials, spectral information that could only be obtained with optical precision instruments in the past can be obtained on a square-inch chip.

From the laboratory to the field, the road to civilian use of a spectrum chip complements AI. This article will talk about how to let spectrum technology enter the consumer market and what AI does to a chip.

It has never been easy to bring scientific research-level technology to the civilian consumer market. Previously, spectroscopy technology was a "flower of high altitude" in the laboratory and has always been blank in the civilian market.

There are two reasons that limit the civilian potential of spectrum technology:

First, it is unaffordable. The collection of traditional spectral information requires the use of spectroscopic instruments, which are bulky and expensive, costing tens to hundreds of thousands. Not only are ordinary families unable to afford them, but also low-profit industries such as agriculture can only stay away.

The second is that it is not used well. Spectral analysis was previously mainly used in laboratories where the research scene was very clean and the signal-to-noise ratio was very low. However, once it is used in the civilian market and is in an open environment, it will face a lot of noise and the signal-to-noise ratio is not good, which directly affects the analysis efficiency and accuracy of spectral information. Therefore, in civilian scenarios, the physical solutions used in the laboratory do not have very good results.

In the AI ​​era, AI technology and spectrum chips complement each other, finally making it possible for spectrum technology and the civilian market to meet.

On the one hand, the hardware is responsible for modulation and AI is responsible for demodulation, which solves the problem of "poor use" of traditional spectral analysis.

The bottom layer of the spectrum chip is CMOS. With the help of spectral modulation technology, the spectrum of the incident light field can be understood, and then the AI ​​algorithm performs data processing and analysis, and automatically extracts features from the spectrum data to remove various noise and perform baseline correction. Therefore, the image information and spectral information obtained by consumer-grade equipment are actually "calculated" through AI.

On the other hand, the physical information of the spectrum + the digital information of machine vision are jointly fed to AI, solving the problem of "affordability" of industry intelligence.

The performance of AI algorithms depends very much on the scale and quality of data. However, many industries have no information or weak information, so they have to use scale effects to improve performance.High model performance results in high computing power requirements and high costs. Spectral information has the characteristics of non-destructive detection, high efficiency and accuracy. With spectral information, the model can be simplified and the complexity of the model can be reduced. Naturally, excessive computing power is not required, thereby reducing the difficulty of application deployment.

It can be said that the interaction between AI technology and spectrum technology has paved the way for the widespread application of spectrum chips in the consumer market.

Spectrum chips have a wide range of application scenarios. Among them, the consumer electronics field has a large audience and high commercial value, and is undoubtedly one of the fields with the greatest commercial potential. At present, AI-powered spectrum chips have explored many application scenarios in the consumer electronics market:

1. High-end mobile phones. High-end flagship mobile phones pay special attention to imaging functions. Spectral chips can emit light of multiple colors, providing richer color choices and higher color reproduction in the lighting and display fields.

2. Smart home. Spectrum chips are gradually being used in smart home products, including refrigerators, sweeping robots, smart door locks, smart cameras, etc. Among them, refrigerators can automatically detect the freshness of ingredients based on spectrum cameras and AI models, and remind users to replace ingredients; projectors use spectrum cameras and color correction algorithms to solve the problem of color distortion of projection images caused by wall color; in the field of smart security, use Spectral chip quickly identifies and tracks target objects , because the spectral information collected is pure physical information, it is not easily deceived by fake hoods and masks, creating a more secure smart door lock; the sweeping robot uses spectral chips + identification algorithms to accurately determine stains, floor Material, intelligent adjustment of cleaning strategy...

3. Automotive electronics. Automobile intelligence is in the ascendant, among which smart cockpits and smart driving are the technological focus of competition among car companies. Spectral chips can play a role in both areas.

In the smart cockpit, the spectrum camera can sense the driver's safety and health status and improve the safety in the car; in smart driving, the spectrum information can accurately reflect the characteristics of road obstacles, such as foam, stones, Roadblocks, etc., have completely different judgment strategies, helping the intelligent driving system to make different judgment strategies.

4. Low-altitude economy. AI spectrum camera + drone can jointly obtain more accurate color information during low-altitude photography and photography, bringing better shooting effects.

5. Others. AI spectrum chips can also be combined with wearable devices to achieve skin health monitoring...

Spectral chips not only have huge potential in the consumer electronics market, but as the digital transformation of all walks of life continues to deepen, spectral information has become a This highly reliable data type can accurately reflect some key indicators of the physical world, and its application value in the B-side industry market is also increasing.

Some organizations predict that the scale of the spectrum chip industry is about tens of billions to hundreds of billions. However, there is a long cycle and uncertainty in the process from design, tape-out to mass production and final commercialization. What tests will an AI spectrum chip go through?

Jilin QiushiThe person in charge of Spectrum told us that they explored the technological innovation of "OCF spectral modulation + algorithm demodulation" in 2017. In 2019, the first Spectrum chip was successfully taped out, and now it has cooperated with leading domestic mobile phone manufacturers and home appliance companies. Cooperation was reached. The combination with AI is also a key layout direction of the company.

The first problem in tempering an AI spectrum chip is funding.

The chip and semiconductor field is a typical resource-intensive industry that relies heavily on funds and talents, and the risks of innovation are high. Therefore, most private capital does not dare to invest. In the past, it mainly relied on government subsidies and project systems. After the establishment of Qiushi Spectrum, it has successively received investment from industrial funds such as China Science and Technology Star, Jilin Science and Technology Investment, Changxing Fund, etc., so that start-ups do not have to worry about funds.

Then comes the data.

Spectral data originates from the physical world, which is difficult and costly to collect. The existing open source data sets in the industry are simply not enough, and most of them have to be collected from scratch. Currently, in the process of developing spectral chips and spectral technology industry solutions, Qiushi Spectrum has gradually collected a large amount of spectral data for model training based on project needs. In the future, it will be necessary to further collect data on southern regions, extreme environments, sand and dust weather, etc. It is expected that the data scale will be 4-5 times that of the current one.

As the data scale increases, computing power challenges follow.

In the past, the company used overseas N cards it purchased to build its own computing power, which was expensive and difficult to maintain. It encountered problems such as small video memory and missing graphics cards that could not be solved, which greatly affected R&D efficiency. Later, the local digital infrastructure was upgraded, and the province's first AI computing center was established. After accessing the computing power service of Changchun Computing Power Center, there is no need to queue for multi-task parallel development. In the past, it took 2 to 3 days to tune a model on the N card. Now, if it is installed that night, the results can be returned the next morning, which has greatly improved R&D efficiency.

However, the manufacturing process does not pose much of a challenge. According to the person in charge, spectrum chips are mainly concentrated in mature processes such as 28nm, 40nm, and 55nm. Domestic manufacturing capabilities can support the large-scale mass production of spectrum chips. Therefore, the spectrum chip was successfully taped out in 2019 and soon entered the commercial stage. Now the public can already experience the more realistic imaging effects brought by the spectrum chip in the mobile phones of leading mobile phone manufacturers.

From this perspective, behind a commercially available AI spectrum chip is the innovation of science and technology investment models, the continuous strengthening of semiconductor and optical industry capabilities, and the upgrading of digital infrastructure. Breakthroughs in hard technology are never achieved overnight, but develop over time.

When the glow of AI shines on a spectrum chip, there are countless possibilities for the integration of the digital world and the physical world.