- China’s new AI chip breaks past traditional computing limits, powering next-gen touchscreens, flight systems, and aircraft navigation.
China has started mass production of the world’s first non-binary chips, adding this new technology to important industries like aviation and manufacturing.
Spearheaded by Professor Li Hongge and his team at Beihang University in Beijing, this project resolves key problems in older systems by blending binary logic with random or probability-based logic. In doing so, it has enabled unprecedented fault tolerance and power efficiency, while smoothly sidestepping US chip restrictions.
Exploring the challenges
Today’s chip technologies face two insurmountable challenges – the power wall and the architectural wall, according to Professor Li. They use too much power, and new chips struggle to work with older systems.
Having been on the search for a solution since 2022, his team came up with a new system called Hybrid Stochastic Number (HSN), which mixes regular binary numbers with probability-based numbers to improve performance.
Binary logic, used by all computers worldwide, uses variables in 0s and 1s to carry out arithmetic operations. However, large-scale binary computations require advanced hardware resources.
In contrast, probabilistic computing uses high voltage signals – how they appear over a set time to represent different values. This method uses less hardware and has already been used in areas such as image processing, neural networks, and deep learning. However, there’s one drawback – it takes longer to process information given the way it represents values.
A solution to the problem
Based on probabilistic computation, Professor Li’s team developed a new smart chip for touch and display in 2023 using leading Chinese chipmaker Semiconductor Manufacturing International Corporation’s mature 110-nanometer process technology.
The project results were published in the IEEE Journal of Solid-State Circuits two years ago. Followed by that, the team came up with another chip for machine learning, fabricated using a standard 28 nm CMOS process.
Apart from HSN, it also features in-memory computing algorithms that reduce the need to move data constantly between the memory and processors. This helps save energy and makes the chip more efficient.
The chip also uses a system-on-chip(SoC) design that combines different computing units to handle multiple tasks simultaneously, unlike traditional chips that process one task at a time.
The chip is now being used in smart control systems, such as touch screens, where it filters background noise to detect weaker signals and improve how users interact with devices.
Apart from HSN, it also features in-memory computing algorithms that reduce the need to move data constantly between the memory and processors. This helps save energy and makes the chip more efficient.
The chip also uses a system-on-chip(SoC) design that combines different computing units to handle multiple tasks simultaneously, unlike traditional chips that process one task at a time.
The chip is now being used in smart control systems, such as touch screens, where it filters background noise to detect weaker signals and improve how users interact with devices.
Plans for the future
Professor Li also told Guangming Daily that his team is developing a special set of instructions and chip design tailored for hybrid probabilistic computing. They plan to use the chip in areas like speech and image processing, speeding up large AI models, and handling other complex tasks.
“The current chip already achieves on-chip computing latency at the microsecond level, striking a balance between high-performance hardware acceleration and flexible software programmability,” Li said.
While China’s move towards non-binary hybrid AI chips is certainly exciting and innovative, it’s important not to overhype the breakthrough yet, as several hurdles still need to be crossed, such as compatibility limitations and long-term uncertainties related to the chip’s usage.
