新型芯片算力可超顶级GPU千倍

Core Insights - A research team from Peking University has developed a high-precision, scalable analog matrix computing chip based on resistive random-access memory (ReRAM), achieving analog computing precision comparable to digital systems [1][2] - The chip significantly enhances computational throughput and energy efficiency, with improvements ranging from 100 to 1000 times over current top digital processors (GPUs) [1][2] Group 1 - The chip addresses complex matrix equation solving, which is essential for applications like communication base station signal processing and AI model training [1] - The research team utilized a novel approach combining new information devices, original circuits, and classical algorithms to create a full analog matrix equation solver with 24-bit fixed-point precision [1][2] - The team successfully demonstrated a relative error as low as 10^-7 after 10 iterations for a 16x16 matrix inversion, showcasing the chip's high precision [2] Group 2 - In terms of performance, the chip surpasses high-end GPU single-core performance when solving 32x32 matrix inversion problems, achieving over 1000 times the throughput of traditional digital processors for 128x128 matrices [2] - The chip's energy efficiency is over 100 times better than traditional digital processors, making it a critical technology for high-efficiency computing centers [2] - The application of this technology in large-scale MIMO signal detection demonstrated high fidelity in image recovery with a bit error rate comparable to 32-bit digital calculations, highlighting its potential in real-time signal processing [2] Group 3 - The breakthrough in analog computing is expected to reshape the computational landscape, providing a promising path for enhancing computational power and potentially breaking the long-standing dominance of digital computing [3]