You may not even realize it, but artificial intelligence (AI) is everywhere nowadays – from smartphone “personal assistants” to games to image upscaling and more. Lots of time, effort, and money have gone into developing chips that can perform AI algorithms faster and more efficiently. The thing is, it can take years to design a chip, which is far too long in the universe of machine learning algorithms. That’s why Google is now using AI algorithms to create AI chips instead of humans.
Azalia Mirhoseini, a Google senior research scientist, said:
We have already seen that there are algorithms or neural network architectures that… don’t perform as well on existing generations of accelerators, because the accelerators were designed like two years ago, and back then these neural nets didn’t exist. If we reduce the design cycle, we can bridge the gap.
Mirhoseini, along with senior software engineer Anna Goldie, developed a neural network that learns to do a subset of chip design – something known as placement optimization. The task is extremely time-consuming for humans because it requires someone to place blocks of memory and logic (or clusters of them called macros) in a strategic way that maximizes power and performance while making the most of the available chip area. At the same time, this has to be accomplished by obeying rules about the density of interconnects.
The procedure requires several weeks-worth of design effort and iteration by human experts (engineers) to map out the ideal placement and produce an acceptable design. To compare, it takes Google’s neural network less than 24 hours to create a design for a Google Tensor Processing Unit after studying chip designs for a while.
The type of AI Mirhoseini and Goldie are using is called a “reinforcement learning” system. It learns by doing, and the more designs it does, the better it gets at it. This model is different than a typical deep learning system that gets trained on large sets of labeled data. A reinforcement learning system can adjust the parameters in their network according to reward signals (in this case, a proxy measure of a combination of power reduction, performance improvement, and area reduction) when they succeed. In this way, the placement-bot becomes better at its task, the more designs it does.
Google’s researchers explain in a paper describing the work posted to Arxiv:
We believe that it is AI itself that will provide the means to shorten the chip design cycle, creating a symbiotic relationship between hardware and AI, with each fueling advances in the other.
We hope AI systems like this will lead to the design of more chips in the same time period, and also chips that run faster, use less power, cost less to build, and use less area.
Such AI systems will inevitably lead to AI becoming even more widely used.
Risto Miikkulainen, a computer scientist from the University of Texas, who was not involved in the work, said:
While most people were taking baby steps, [the researchers] took a giant leap into the unknown. This is one of those papers that could launch a lot of future research.
At the moment, the system is only capable of producing simple AI systems. Still, the researchers continue working on improving it so that eventually, it’ll be able to spit out algorithms unimaginable by human programmers.
The post AI Designs AI Processors In 24 Hours While Humans Take Weeks appeared first on Intelligent Living.
