Real-time reconstruction of the brain's complex cortical folds requires computational power that traditional hardware struggles to provide without critical latency. A team led by Yuchao Yang at Peking University, in collaboration with the Shanghai Institute of Microsystem and Information Technology, has broken this barrier by developing a phase-change memristor chip capable of completing these operations in under 10 milliseconds, matching the operational pace of the human brain.

Psst!Add AlexTech.ai asPreferred Source on Google

Solving the Memory Bottleneck

The primary limitation of conventional computing is the constant data transfer between the processor and memory, which creates delays and high energy consumption. The new chip addresses this by implementing in-memory computing, performing calculations directly where the data is stored. This architecture eliminates unnecessary data movement, drastically reducing energy requirements and accelerating system response times.

NVIDIA H100 Compared to A100 for Training GPT Large Language Models ... — https://www.techpowerup.com/307922/nvidia-h100-compared-to-a100-for-training-gpt-large-language-models

Outperforming NVIDIA A100 GPUs

Manufactured using a 40-nanometer process, the chip occupies only 0.28 square millimeters and operates at 50 MHz. Findings published in Science show a significant performance leap: the device is 3.82 to 36.27 times faster than state-of-the-art ASICs and consumes 11.75 to 24.73 times less power. Specifically for cortical surface reconstruction, the chip achieved a speedup of up to 478 times compared to an NVIDIA A100 GPU.

Toward Digital Twins and Real-Time Surgery

The ability to generate smooth, topologically consistent 3D surface meshes in milliseconds opens immediate application scenarios. Beyond enhancing brain-computer interfaces, this technology could enable real-time surgical navigation and instantaneous screening for neurodegenerative diseases such as Alzheimer's and Parkinson's, shifting brain modeling from a slow offline process to an immediate diagnostic tool.

NVIDIA A100 平替——NVIDIA A800 - 知乎 — https://zhuanlan.zhihu.com/p/573471332

The Future of Neuromorphic AI

This memristor-based approach is part of a broader trend toward neuromorphic computing, aiming to overcome the energy bottlenecks of current AI models. While some research explores biological neurons on silicon, integrating memristors into traditional semiconductor fabrication offers a more scalable path to achieving brain-like efficiency and plasticity without the ethical complexities of biocomputing.