The Realm of 'Superintelligence'... AI Designing Semiconductors [Tech Talk]

Semiconductor Design Beyond Human Comprehension
AI Drawing Schematics Much More Efficiently Than Humans
Is an Era of 'Fabless' Business Automation Coming?

A semiconductor company that owns only the design blueprints without a factory and entrusts manufacturing to other companies is called a 'Fabless' company. The name comes from the fact that it lacks a 'fab,' which means a state-of-the-art semiconductor factory.

Although fabless companies do not have factories, they are instead armed with countless intellectual properties (IP) and engineers that ordinary manufacturers cannot even imagine. Thanks to this, they have been able to negotiate semiconductor production prices in balance with specialized contract manufacturers such as TSMC and Samsung Foundry.

However, the recent advancement of artificial intelligence (AI) might bring a seismic shift to the fabless business model. AI exhibiting 'superintelligence' that surpasses the human brain could potentially automate the chip design process.

Semiconductor Complexity Beyond Human Understanding Becomes a Weapon Against Theft

Google's ARM architecture-based Axion CPU. This computer chip was designed by DeepMind, Google's AI research subsidiary, using 'semiconductor design artificial intelligence.'
[Image source=Google]

The fabless companies’ weapon lies in their IP and powerful research and development (R&D) capabilities. If semiconductor design were just simple blueprints, coexistence between foundries (contract manufacturers) and fabless companies would not be possible. Foundries could have 'reverse-engineered' fabless designs and produced similar products at any time.

However, the complexity of advanced semiconductors has now exceeded the scope of human understanding. The amount of work required to design a single logic chip is beyond imagination, making it impossible to reverse-engineer a product simply by 'looking' at the design. In other words, the inscrutability of semiconductors itself serves as a shield for fabless companies.

For this reason, semiconductor design demands excellent semiconductor engineers, R&D, and 'time.' Especially, finding engineers capable of creating complex logic chips such as central processing units (CPUs) is known to be very difficult. Currently, only a handful of countries such as the United States, the United Kingdom, and Israel possess the capability to create top-level CPU designs. Even South Korea’s largest semiconductor company, Samsung Electronics, has established separate R&D centers in Austin, Texas, and California to research logic chips.

'Superhuman' AI Drawing Chip Blueprints Surpasses Human Engineers

Is it possible to automate semiconductor design even a little? Recently, DeepMind, an AI research lab under Google, released a new AI model called 'AlphaChip.' AlphaChip is the world’s first 'AI that designs chips.' Like DeepMind’s previous AI models such as AlphaGo and Gemini, it is a hybrid model combining deep learning and reinforcement learning, and it improves performance by learning from various data.

Google's semiconductor design AI AlphaChip performing the semiconductor 'placement & routing' task. This design process optimizes the arrangement of device blocks to increase the density of semiconductors. [Image source=Google]

Amazingly, AlphaChip is already being used to manufacture chips within Google. AlphaChip contributed to the design of Google’s AI accelerator 'TPU,' which is considered a rival to Nvidia’s graphics processing units (GPUs). It also helped create Google’s own CPU, Axion. Axion is a chip based on ARM’s server CPU core called 'Neoverse' and already utilizes ARM’s automated design tool, Compute Subsystem (CSS).

In other words, Google’s semiconductor design process, unlike previous semiconductor designs, has benefited from various automation programs. This is likely the secret behind how Google, a giant IT company with little prior involvement in semiconductors, was able to establish its own chip design team so quickly.

Most importantly, DeepMind announced that AlphaChip’s design optimization feature called 'Superhuman Layout' has already outperformed skilled human engineers. The automated section by Superhuman Layout is 'Placement & Routing (P&R),' which is the process of arranging wires that connect each semiconductor element to create the design blueprint.

TPU data center, the hardware brain of DeepMind. The design of the TPU is also carried out by the AlphaChip.
[Image source=DeepMind homepage]

In the case of the 6th generation TPU called 'Trillium,' AlphaChip completed the work using 6.2% less total wire length than human experts. A reduction in wire length means there is less empty space between elements, which implies that more elements fit inside a single chip, resulting in a stronger and more efficient chip design. Above all, AlphaChip’s P&R skills are increasingly widening the gap with humans. At the current pace, AlphaChip’s P&R efficiency will soon surpass humans by a 10% margin.

Will AI Become Fabless in the Near Future?

Layout is only a very small part of the chip design process, and there is still a long way to go before automation can completely replace human GPU and CPU engineers. However, the rapid improvement of AI engineers’ skills and the fact that AI is already used in designing commercial products are significant.

As mentioned earlier, fabless companies have been creating chips through massive R&D, time, and engineering manpower. If a large portion of chip design becomes automated and the required personnel decreases, the barriers for a company to become fabless will be lowered accordingly. Also, the time required to create the latest generation computer chips will be reduced.

Of course, even if these changes materialize, the market dominance of fabless companies will likely not diminish. Instead, the industry’s winners and losers may be determined by who owns the 'superhuman AI' capable of creating future semiconductors beyond our understanding.