This tiny chip will break Moore's law to become a leader in artificial intelligence

The Eyeriss chip appeared before the big "big hands" of the world technology industry, then left a deep impression on the possible future of artificial intelligence.

The early morning sunshine over Palm Springs, California, a wonderful setting for Vivienne Sze to stand up on the small stage, make the presentation perhaps the most difficult career. Sze is well versed in every industry in the presentation: she is about to talk about new chips , being developed at her lab at the Massachusetts Institute of Technology (MIT).

The tiny chips carry the promise of an artificial intelligence system that can be installed on any device - small technological toys that are not capable of containing the enormous power of an AI system. . Until now, AI-based calculations were limited to giant data centers.

Confident that, Vivienne Sze is still . standing in front of the event scene, and before the big minds of the world are scattered there.

This is MARS , a conference dedicated to guests; The luxurious resort setting looks like a straightforward sci-fi movie: the ground and the air are full of running and flying robots, scientists, science fiction writers I thought I was talking to each other in some supreme language. Only a few researchers were invited to give presentations, answer questions.

The crowd in the hall is a combination of 100 world-class researchers, CEOs and investors. And the organizer, no one else, is the highest executive billionaire Amazon, Mr. Jeff Bezos.

There are speakers at MARS that bring karate robot products, drones that can flap their wings like a real insect, there is also a blueprint of the human Mars base that will how. Ms. Sze's chip looks humble, less flashy like those visual products.

In front of our human eyes, this chip is no different from what is available in smartphones. But it can be affirmed: Ms. Sze's chip is the second most important thing to attend at the MARS event.

The small chip opens up large doors

New chips, like what Sze is developing in the lab, are the future of artificial intelligence; itself is the brain of everything running robots, flying in the room. Until now, AI software mostly runs on graphics chips, so an effective new hardware will make AI algorithms more powerful than ever, even opening up our potential. know.

New AI chips will create a new generation of robots, or simply allow smartphones to create an augmented reality environment with a level of authenticity never before.

It is designed to contain more deep learning algorithms, lines of code that have changed the world. We can even dream about the utopian, when algorithms can 'evolve' to open new doors.

'We need new hardware because Moore's law is preventing technological advances , ' Sze said, referring to the law coined by Intel co-founder Gordon Moore. The talented engineer predicted that the number of transistors on each chip would double in the 18-month period, through which the computer would be even more powerful.

But Moore's law is gradually reaching the physical limit, when we point down to the atomic size as the end of the rung. This is why we need a new, more efficient computer structure to prepare for a long journey of development.

We need a new generation of artificial intelligence chips , and chip makers have a different task: they need to make sure the United States holds the leading position, who has the most powerful chip in the world, it must have a computer. the most powerful in the world. Ms. Sze's chip is receiving investment from the Advanced Defense Research Projects Agency (DARPA), which is looking forward to a new chip design.

The need for a new chip comes from the rise of deep learning - a new way that allows a machine to perform extremely difficult and extremely useful tasks. Instead of giving the computer a set of rules to follow as before, now computers can program themselves to act (understand that way). People put a huge amount of data into a neural network system (with mechanisms similar to the human brain), editing the input to get the right results.

When 'training' an AI long enough, it can track traces, difficult to detect patterns in a large database. This technique appears everywhere in the modern world: from facial recognition to predicting disease from input images, from describing disease symptoms.

Technology generates a new "chip match"

Moore's law is not limited to deep learning. A neural network can allow mathematical calculations to run in parallel with each other, which is why the on-card graphics card effectively handles deep learning algorithms. However, a chip specialized for deep learning calculations would have been more effective.

The emergence of AI has led the world chip industry to launch a new race, which has not appeared for several decades. The market appears new big men with great potentials never seen before; According to MIT Technology Review, China has never had a true chip industry, and AI will be a perfect start.

Large technology companies obviously want to contribute and commercialize the new AI industry; Google, Microsoft and Amazon are all developing their own deep learning chips . Small companies also don't want to be out of the race.

Contrary to what we thought, Ms. Sze asserted that the chance to get out really wasn't designed, so the chip deep learning was as powerful as possible. The chip knows regulation, energy efficiency is important, because it will need to go beyond the framework of large data centers, use electricity itself, the devices that install them to operate.

The term industry calls this a "edge" operation.

' AI will be ubiquitous, optimizing energy when operating AI will be an important factor in the future, ' said Naveen Rao, vice president of artificial intelligence products at Intel.

For example, immediately with Ms. Sze's chip: it uses energy more efficiently by special design, reducing bottlenecks between data storage and data analysis; Besides, hip also takes advantage of old data skillfully. Prior to arriving in MIT, Ms. Sze worked at Texas Instruments, who was a pioneer in revolutionizing video compression, allowing it to be more effective.

In the field of continuous transformation such as deep learning, the challenge for developers is to ensure the chip performs all the tasks in modern times. It's easy to create a powerful chip that does a certain task, but such a product will soon be obsolete, as the trend quickly turns itself into a new one.

Ms. Sze's chip is called Eyeriss , developed in partnership with Joel Emer, a researcher at Nvidia and a professor at MIT. Eyeriss is tested in parallel with a number of other standard processors, to compare the operating speed of different deep learning algorithms. Balancing efficiency with flexibility, Eyeriss's performance is 10, even 1,000 times the current hardware. Research results on Eyeriss were published last year.

Simple AI chips could shake the research industry. The leading smartphones of the present time have used optimized chips for deep learning algorithms, used in image recognition and voice recognition. More efficient chips will allow those devices to run more complex algorithms, bringing new and more efficient applications.

Not to mention the future of vehicles - self-driving cars also use powerful AI chips. Current self-driving cars must use 'computer stalls' in the trunk. Imagine the ability of smart cars to be fitted with advanced AI systems, like a growing iron tiger with jet wings.

Vice president Rao said the MIT chip has potential, but in order for the chip to truly enter life, many other factors must also succeed, not just its hardware structure. One of the most important points is developing software applications on hardware.

Ms. Sze's lab is also looking for ways to design software, based on existing knowledge with current computer chips. This aspect has far surpassed the concept of 'deep learning'.

Together with Sertac Karaman from MIT's Department of Aeronautics and Astronautics, Ms. Sze developed another low-power chip, called Navion . It has the ability to create extremely effective 3D maps and navigation, which can be mounted for drones. Special point of the research process: they make chips based on available navigation algorithms. The Navion chip is a product that shows AI hardware and software combined, giving us the first products of the new generation of technology.

Ms. Sze's chip is not as eye-catching as a butterfly-like drone, but is still honored as the second most important technology ever brought to the MARS event. Sze said many speakers at MARS were immensely interested in the power in the tiny chip.'People have found many different applications for chips , ' she said.

In subsequent technology shows, it is possible that in flashy robots, the power of Eyeriss, the "prophetic" chip is a pioneer in the era of artificial intelligence.