Samsung uses artificial intelligence to automate the complex and delicate process of creating cutting-edge computer chips.
South Korean giant Samsung is among the first to use AI in creating its chips. Synopsys is a top chip design software company that Samsung uses to create its chips. Aart de Geus (chairman and coCEO of Synopsys) says that this is the first commercial processor design using AI.
Nvidia and Google have also discussed AI-based chips. The Synopsys tool DSO.ai may be the most useful, as it works with many companies. According to industry watchers, the tool could accelerate semiconductor development and unlock new chip designs.
Synopsys also has a valuable resource for creating AI-designed chips: years worth of semiconductor designs that can be used in training an AI algorithm.
Samsung's spokesperson confirmed that Exynos chips are being designed using Synopsys AI software. This software is used in smartphones including Samsung's own handsets. Samsung's newest smartphone, the Galaxy Z Fold3, was unveiled earlier this week. Samsung did not reveal whether the AI-designed chips were yet in production or which products they might appear in.
AI is changing how chips are made across the board.
Advertisement
In June, Google published a research paper that described how it used AI to arrange components on Tensor chips. This is the same AI technology that it uses in order to train and run AI programs within its data centers. The next Google smartphone, the Pixel 6, will include a custom-made chip made by Samsung. Google declined to confirm whether AI was involved in the design of the smartphone's chip.
Nvidia and IBM, two of the world's largest chipmakers, are also exploring AI-driven chip design. Cadence, which is a competitor of Synopsys' chip-design software, also develops AI tools that help with the mapping out of blueprints for new chips.
Mike Demler, a Linley Group senior analyst who tracks chip design software says that artificial intelligence is well-suited for arranging billions upon billions of transistors on a chip. He says that artificial intelligence is well-suited to solving complex problems. It will become an integral part of the computing toolkit.
Demler said that AI is often expensive because it requires large amounts of cloud computing power to train an algorithm. He expects AI to become more affordable as computing costs drop and models become more efficient. He also said that not all tasks in chip design can be automated. Expert designers are still required.
Modern microprocessors have many components that must be combined in a way that is efficient. It takes a lot of time and decades to design a new chip. Chip designers who are skilled in designing chips have a deep understanding of how decisions affect each stage of the process. This understanding is not easily captured in computer code. However, machine learning can capture some of the same skills.
Synopsys uses an AI approach, along with Google, Nvidia and IBM. It employs a machine learning technique called reinforcement to design a chip. Reinforcement learning is a method of training an algorithm to do a task using reward or punishment. It has been proven effective in capturing subtle and difficult-to-codify judgments.
Advertisement
This method automatically draws up basic design elements, such as the placement of the components and how they should be wired together. It also allows you to simulate different designs and learn which one produces the best results. This software can be used to speed up the design process and allows engineers to explore new designs faster. Synopsys reported that a North American manufacturer of integrated Circuits had seen a 15% improvement in the performance of their chips by using the software in a blog post.
DeepMind, a Google subsidiary in 2016, used reinforcement learning to create AlphaGo. This program was capable of mastering Go and defeating a world-class Go player.
De Geus claims that his company discovered reinforcement learning could be used for chip design. De Geus states that the company was able to achieve the same results in just a few short weeks as a team with experts a year ago. HotChips will host de Geus' presentation on the technology and its progress at 23 August, a conference in semiconductor technology.
Stelios Diantis, Synopsys' senior director for artificial intelligence solutions, said that the DSO.ai software could be set up to prioritize different goals such as energy efficiency or performance.
Semiconductors and the tools that make them are becoming increasingly valuable assets. The US government sought to limit the supply of chipmaking technology in China, a key competitor, and some politicians called for software to be added as an export control item.
AI-designed chips are also a possibility. This allows for AI to be used to tailor software to run faster on a chip. These neural network algorithms, which are often used in modern AI, run on specialized AI chips.
AI-powered codesign of hardware and software is a growing trend, according to Song Han, a professor at MIT who specializes on AI chip design. We've seen promising results.
This story first appeared on wired.com
