Researchers from Nanjing University’s Brain-Inspired Intelligence Technology Center have developed an analog compute-in-memory chip that achieves a record root-mean-square error of 0.101% in vector-matrix multiplication. The study, published in Science Advances, highlights a fundamental shift from traditional physical parameter dependence to device geometric size ratios for computational weights, ensuring enhanced stability.
The chip maintains high precision across extreme temperatures from -78.5°C to 180°C, with errors between 0.130% and 0.155%, and exhibits minimal output current variation (under 0.21%) in 10T strong magnetic fields. In real-world tests, it reached 97.97% accuracy on the MNIST dataset for neural network inference—matching 64-bit floating-point software results—and demonstrated strong performance in solving fluid dynamics equations.
ICgoodFind : This innovation provides a robust hardware solution for edge computing and intelligent processing in harsh environments, underscoring analog computing's advantages in efficiency and reliability.
