Ultrahigh-precision analog computing using memory-switching geometric ratio of transistors.

Area of Science:

• Materials Science
• Computer Engineering
• Artificial Intelligence Hardware

Background:

• Traditional analog computing relies on physical quantities prone to fluctuations, limiting precision.
• Environmental changes and programming can compromise analog system accuracy.
• Need for stable and precise analog computation in AI hardware.

Purpose of the Study:

• To develop an ultrahigh-precision analog computing chip.
• To shift reliance from fluctuating physical quantities to stable geometric features.
• To demonstrate a novel analog in-memory computing approach.

Main Methods:

• Utilized the geometric ratio of transistors instead of intrinsic physical quantities for computation.
• Developed an analog in-memory computing chip using a standard complementary metal-oxide semiconductor process.
• Implemented a weight remapping technique to enhance computing accuracy.

Main Results:

• Achieved the highest precision reported to date in analog computing.
• Demonstrated ultrahigh computing accuracy with a root mean square error of 0.101% in parallel vector-by-matrix multiplication.
• Maintained high precision at extreme temperatures (-78.5°C and 180°C) with errors of 0.155% and 0.130%.

Conclusions:

• Leveraging stable geometric features of devices significantly enhances analog computing precision.
• The developed chip offers a promising solution for stable and accurate AI hardware.
• This work advances the field of analog computing by pushing precision boundaries.