Achievement by Assistant Professor Longyang Lin's Team at Our School Selected as One of SUSTech's Top 10 Scientific and Technological Advances of 2025!

2026-01-23 General News views：9

Top 10 Scientific and Technological Advances of 2025 Released!

These Top 10 Scientific and Technological Advances represent the annual outstanding achievements selected across the entire university, including departmental recommendations, expert reviews, online voting by faculty and students, and deliberation by the review committee office. Ten representative annual research achievements were ultimately determined. The selected achievements are primarily distributed across research fields such as physics, chemistry, earth sciences, environment, semiconductors, life sciences, and medicine. The university is also the first affiliated institution and corresponding unit for these achievements.

The team of Assistant Professor Longyang Lin from the School of Microelectronics, was selected for: Breaking the "Memory Wall" Bottleneck: A High-Efficiency All-Digital Non-Volatile Compute-In-Memory AI Chip.

Breaking the "Memory Wall" Bottleneck: A High-Efficiency All-Digital Non-Volatile Compute-In-Memory AI Chip

• Recommending Unit: College of Engineering

• Achievement Team: Longyang Lin's Team

• Affiliated Unit: School of Microelectronics

  

Achievement Summary:

• Research Background: In the post-Moore era, the demand for computing power and energy efficiency for AI applications is growing explosively. However, the "memory wall" and "energy efficiency wall" issues caused by the separation of computing and memory in the traditional von Neumann architecture are becoming increasingly prominent. Compute-In-Memory (CIM) technology is regarded as a key path to break through this bottleneck. However, traditional analog CIM solutions face core challenges such as limited computational accuracy, weak noise immunity, and difficulties with process scaling.

• Researchers and Research Content: The team of Longyang Lin from the College of Engineering's School of Microelectronics, has proposed and tape-out validated the world's first all-digital non-volatile compute-in-memory (nvDCIM) architecture based on 40nm CMOS technology integrated with Magnetic Random-Access Memory (MRAM). The team innovatively invented an "in-cell multiplication and digitization" technique, achieving lossless digital computing in memory. They also designed a reconfigurable accumulator supporting multi-bit precision and an all-digital readout circuit, completely eliminating the need for power-hungry and area-intensive analog-to-digital converters (ADCs). Furthermore, the team proposed a "toggle-rate-aware" software-hardware co-training algorithm that significantly reduces the chip's dynamic power consumption during inference execution without degrading model task accuracy. The related results were published in Nature Electronics on October 16, 2025.

Recommendation Justification:

This research overcomes the critical challenge of achieving high-precision, high-efficiency computing with non-volatile memory in advanced process nodes, providing a new technological pathway for the energy-efficient deployment of edge AI chips. It was also selected by the journal's editors as a Research Briefing for key promotion.