AI makes chips its own way

AI has the potential to change the world that we live in completely. One of those major changes that AI is forecast to alter dramatically is in the jobs we do. It’s a subject I try to follow closely online, especially the social aspects. Many people think the jobs that we will lose will mainly be low-level repetitive jobs, but relatively simple automation is doing that already. Complex AI is not required to pack crates or move them around a warehouse. There may be some AI aspects in the role, but that is nothing compared to driving a vehicle safely. Logistics companies like Amazon, along with automation companies have been working on automating tasks like these for a number of years. AI could take things much further – and not only do the jobs themselves, but change the actual way we do things.

Examples of the type of role that AI could potentially perform are widely available online. For example, there have been many stories about how AI has been trained to identify cancers from MRI imaging more accurately than medical professionals. Also, docks in Singapore and mines in Sweden have already stopped using human drivers and all transport is AI controlled. Today I read another article that brought things a lot closer to home in our industry.

Google has written a paper on a way for its AI to plan the layout of ICs, one of the most complex tasks around. Normally, chip floorplanning requires months of effort by physical design engineers and has proven very difficult to automate effectively. In the new paper, the company presented a deep reinforcement learning approach to chip floorplanning. It claims that its method can automatically generate floorplans in under six hours that are comparable or superior to those of human design teams in power, performance and area. To get the edge in performance, Google posed chip floorplanning as a reinforcement learning problem, and developed an edge-based graph CNN architecture capable of learning rich and transferable representations of the chip. The method uses past experience to become better and faster at solving new instances of the problem. This allows chip design to be performed by AI with more experience than any human designer. It is not just a theoretical idea either, Google used the AI to assist in the design of the company’s next generation of AI accelerators. The paper also states that more powerful AI-designed hardware will help design better chips, which will then replace the AI hardware to design even better chips.

However, the results with AI are often not what you would expect. Dr. Adrian Thompson, a researcher from the Department of Informatics at the University of Sussex used AI to design a circuit on a small FPGA in the 1990s. After thousands of iterations, the FPGA design worked as intended. However, when Thompson tried to see how the design worked, he was left confused as there were feedback loops there that apparently did nothing, but when removed, the circuit stopped working entirely. The design would not even work correctly in an identical FPGA.

https://ai.googleblog.com/2020/04/chip-design-with-deep-reinforcement.html

https://www.damninteresting.com/on-the-origin-of-circuits/