Could AI-designed chips herald a revolution in compute?

Could AI-designed chips herald a revolution in compute?

Synopsys company claims that in time, AI-designed chips can help realise processors that have 1,000 times the compute performance of today's processors

Representative image. Credit: iStock photo

Microprocessors have come a long way from the days of the humble Intel 8008. In their nearly 50 years of existence, microprocessors have grown vastly more powerful, efficient, capable of handling even quantum computing simulations, and have greatly contributed to an increased quality of life of people.

However, all this has come at a cost: Over the last 10-15 years, microprocessors have grown exponentially more complex, with the latest CPUs having anywhere up to 20 billion transistors across all their components such as cache, logic and I/O systems.

To help mitigate the ever-increasing complexity of chip design, US-based tech company Synopsys is attempting to pioneer what it calls AI(Artificial Intelligence)-designed chips.

At the 2021 Hot Chips conference, Synopsys announced an expansion of its electronic design technology called DSO.ai (Design Space Optimization AI), which it says is the industry's first autonomous artificial intelligence application for chip design.

DSO.ai is inspired by AlphaZero by Alphabet's (parent company of Google) DeepMind, which was trained to master highly complex games like Chess and Go, and is essentially an AI and reasoning engine capable of searching for optimisation targets in chip design.

The company refers to this technology as a shift from software-defined hardware to software-designed hardware, and the technology has already seen practical applications by companies such as Samsung Electronics and Renesas.

Companies like Google and Nvidia are already in the process of investing in AI to help design some of their chips.

However, even with cutting edge design tools, massive international teams and years of R&D, it can be highly difficult to design a chip as large and complex as any modern processor, and it is known that higher complexity in design also means the chip has a higher chance of having bugs during testing, which can add to the time to market.

This is where Synopsys says its DSO.ai comes in. Using a model called reinforcement learning (the same technology used by Google to win in chess and Go), the company says the technology can help automate the entire design process for chips, from the logic, to the layout and the function of chips, with vast improvements in power over traditional human-designed chips. Currently, DSO.ai already works on the layout portion of the design process, using AI to help determine optimal layout of the chip's physical components.

However, the company does not see AI replacing humans entirely. Speaking to DH, Stelios Diamantidis, Senior Director, Synopsys AI Solutions, said: "AI chip design technology automates some steps in the process that ultimately elevates the work of the human designer.

At first, people were drawing circuits by hand. Then, structured place around systems came on the scene and they started designing modules that could be replicated. After that, synthesis came along and designers started using behavioral models.

In recent years, IP (Intellectual Property) core has become very important. Every major advancement in the industry that achieves a productivity boost elicits these same questions on whether or not humans will become obsolete.

"The only thing that changes is the level of abstraction and the level of direction that the designer plays. History has shown that instead of dwindling in number, designers have become in more high demand, and they produce more chips," added Diamantidis.

Diamantidis also said that AI-designed chips can help reach designer's higher transistor density (the number of transistors fits in one square mm of area for a given manufacturing node) than purely human-designed chips and in a much shorter span of time.

The company claims that in time, AI-designed chips can help realise processors that have 1,000 times the compute performance of today's processors, with vast improvements across every aspect of design, be it the quality of result, cost of result, or time to result.