Peking University's Analog Chip Outperforms Nvidia's H100 in Speed and Energy Efficiency

Researchers at Peking University have developed a high-precision, scalable analog matrix computing chip that significantly outperforms leading digital processors, including Nvidia's H100 GPU, in both speed and energy efficiency. This breakthrough, detailed in a study published in Nature Electronics on October 13, 2025, leverages resistive random-access memory (RRAM) to achieve computing throughput and energy efficiency 100 to 1,000 times greater than current top digital processors when solving complex scientific problems like large-scale multiple-input multiple-output (MIMO) signal detection.

Analog computing processes information through continuous electrical signals, offering potential advantages in speed and energy efficiency over traditional digital computing, which relies on binary code. Historically, analog systems have been limited by issues of precision and scalability, restricting their practical applications. The advent of RRAM technology has opened new avenues for analog computing by enabling in-memory processing, thereby reducing the energy and time costs associated with data transfer between separate memory and processing units.

The Peking University team's chip integrates RRAM cells with novel circuit designs and classical algorithms to achieve 24-bit fixed-point accuracy, a significant improvement over previous analog systems. This precision is achieved through an iterative algorithm that combines low-precision analog matrix inversion with high-precision analog matrix–vector multiplication operations.

In practical applications, the chip demonstrated exceptional performance in solving large-scale MIMO signal detection problems, achieving throughput and energy efficiency 100 to 1,000 times greater than leading digital processors like Nvidia's H100 GPU.

Nvidia's H100 GPU, based on the Hopper architecture, is renowned for its high performance in AI and high-performance computing tasks. It features up to 16,896 CUDA cores and 528 Tensor Cores, delivering significant computational power. However, the Peking University chip's architecture allows it to perform specific tasks, such as large-scale MIMO signal detection, with substantially higher efficiency and speed.

This development addresses longstanding challenges in analog computing, particularly concerning precision and scalability. By achieving accuracy comparable to digital systems, the chip opens new possibilities for energy-efficient computing in fields such as artificial intelligence and next-generation communication systems. The researchers plan to develop larger, more integrated versions of the chip to tackle even more complex computations, potentially revolutionizing the landscape of high-performance computing.

The implications of this breakthrough are far-reaching. In artificial intelligence, the enhanced speed and energy efficiency could lead to more powerful and sustainable AI models. In communication systems, particularly with the advent of 6G, the chip's capabilities could support more efficient signal processing, leading to faster and more reliable networks.

As the Peking University team continues to refine and scale their technology, the computing industry may witness a paradigm shift, with analog computing playing a more prominent role alongside digital systems. This advancement not only showcases China's growing prowess in technological innovation but also sets the stage for a new era in computing efficiency and performance.