Fall 2021: A season of pumpkins, pecan pies, and peachy-colored phones. Every year, Apple, Samsung, Google, and others lose out on their most recent accomplishments. These machines in the consumer professional calendar no longer impress with the surprise and wonder of those early days. But behind the scenes, there is a strange thing happening.
Google’s latest offering, the Pixel 6, is the first phone to have a separate AI chip that sits next to its standard processor. And the chip that runs the iPhone has been around for the past few years with what Apple calls a “neural engine,” dedicated to AI. All of the chips are well-suited to the types of computers used to train and deploy machine learning software on our equipment, such as AI that supports your camera. Almost unknowingly, AI has become part of our daily lives. And it is changing the way we think about computers.
What does that mean? Computers haven’t changed much in 40 or 50 years. They’re small and fast, yet boxes with processors that run instructions from people. AI varies in three areas: how computers are designed, how they are designed, and how they are used. Finally, change the status quo.
Pradeep Dubey, director of the Intel Computing Lab at Intel, said: “Computer change is shifting from increasing numbers to decision-making.” Or, as MIT CSAIL director Daniela Rus puts it, AI pulls computers out of their boxes.
Too fast, too slow
The first changes affect how computers, as well as chips we operate. The traditional benefits of computers came as the machines ran faster in computing each other. For decades the world has benefited from the speed of technology that comes consistently as manufacturers in accordance with Moore’s Law.
But the types of in-depth learning that use AI software here in operation require a different approach: it requires more numbers to be implemented simultaneously. This means that a new type of chip is needed: one that can move data faster, making it available where it is needed. When in-depth learning exploded on the scene a decade or so ago, there were special computer tools that were very good at this: photo frames, or GPUs, designed to display pixels several times per minute.
Anything can be a computer. Indeed, many household items, ranging from toothbrushes to door-to-door cans, have come to the fore.
Now developers like Intel and Arm and Nvidia, which provide many of the first GPUs, are trying to develop AI-compatible devices. Google and Facebook are also pushing to enter the industry for the first time, in a competition to get AI side by side through tools.
For example, the chip inside the Pixel 6 is a new feature for Google, or TPU. Unlike traditional chips, which focus on ultrafast, real-time counting, TPUs are designed for complete but low-cost computing that requires neural networks. Google has been using the same chips in-house since 2015: it changes images of people and language search queries. Google’s sister company DeepMind uses it to train its AI.
Over the past few years, Google has made TPU available to other companies, and the chips – as well as their counterparts – are becoming more and more consistent within the global database.
AI is also helping to create its own architecture. In 2020, Google adopted an innovative learning-to-type AI system that learns how to solve tasks by testing-optimizing the design of a new TPU. AI finally came up with a new unique design that no one could imagine — but it worked. This type of AI can one day have a good, very effective chip.
Show, don’t say
The second change affects how computers are told what to do. For the past 40 years we have been developing computer software; for the next 40 we will be training them, says Chris Bishop, chief executive of Microsoft Research in the UK.
Traditionally, for a computer to do something like cognitive recognition or visualization, first programs have to create computerized rules.
With the advent of machine learning, programmers no longer write rules. Instead, they create neural networks that study the rules on their own. It’s a completely different way.