Skip to main content Audio 2 ![Image 1: LogoBest Blogs](https://www.bestblogs.dev/ "BestBlogs.dev")

Search Ctrl+K

Change language Switch ThemeSign In

[](https://www.bestblogs.dev/en/explore/brief "Daily Brief")[](https://www.bestblogs.dev/en/explore/newsletter "Weekly Picks")[](https://www.bestblogs.dev/en/explore/topics "Topics")[](https://www.bestblogs.dev/en/worldcup "World Cup Special")

[](https://www.bestblogs.dev/en/settings "Settings")[](https://www.bestblogs.dev/en/docs "Help Center")

Narrow Mode

周末利好！芯片，突传重磅！全球首款问世

A joint team from Peking University and the Chinese Academy of Sciences has developed the world's first memristor-based neurodynamic chip, compressing computational latency to milliseconds for the first time and achieving up to a 478x speedup over GPUs, providing a new hardware foundation for brain-computer interfaces and brain disease treatment. ![Image 2: 券商中国券商中国](https://www.bestblogs.dev/articles?sourceid=01ca8fe0 "View More From This Source")Follow·

Yesterday·1921 words (about 8 min)

·View Source →

AI Summary & Key Points

Summary

The article reports that Yang Yuchao's team from Peking University, in collaboration with the Chinese Academy of Sciences, has successfully developed the world's first neurodynamic chip based on phase-change memristors, with the results published in the journal _Science_. The chip leverages the physical properties of memristors to achieve 'controlled in-memory computing', compressing single-step operation latency to 2.12 milliseconds. In cortical reconstruction tasks, it achieved a 478x speedup compared to the NVIDIA A100 GPU, breaking through a half-century bottleneck in real-time neurodynamic computation. The latter part of the article reviews the latest developments in the brain-computer interface industry, including its first inclusion in a government work report and the introduction of specific policies in multiple regions. Citing data from McKinsey and the China Academy of Information and Communications Technology, it points out that the global market size for medical applications is expected to reach $40 billion by 2030, indicating enormous market potential.

Main Points

* 1. The world's first memristor neurodynamic chip achieves millisecond-level real-time computation, breaking through the traditional bottleneck of separated storage and computing.

The team utilized the conductance drift phenomenon of phase-change memory to encode adaptive step-size search as a physical process, completing calculations within a 0.28 square millimeter in-memory computing array, with a single-step latency of only 2.12 milliseconds.

* 2. The chip's performance vastly surpasses existing GPUs and specialized accelerators, providing a real-time hardware foundation for brain science applications.

In cortical reconstruction tasks, it achieved a 478x speedup over the A100 GPU and reduced power consumption by over 11 times, making applications like personalized brain digital twins and intraoperative neural navigation feasible.

* 3. The brain-computer interface industry is gaining momentum, with broad prospects for medical applications driven by both policy and market forces.

Brain-computer interfaces were included in a government work report for the first time, and multiple regions have introduced specific policies; McKinsey predicts the global market size for medical applications will reach $40 billion by 2030.

周乐 2026-07-04 20:42 广东

!Image 3

券商中国微信公众号：quanshangcn

!Image 4: 图片

芯片领域传来新突破。

据最新消息，北京大学集成电路学院教授杨玉超团队联合中国科学院研发出全球首款忆阻器神经动力学芯片，相关研究成果已经登上《科学》期刊。杨玉超表示，这一突破为脑机接口和脑疾病诊疗开启了全新想象空间。

当前，脑机接口产业热度正持续攀升，相关技术在医疗健康领域的应用规模增长较快。据麦肯锡测算，全球脑机接口在医疗应用领域的2030年市场规模有望达到400亿美元，到2040年突破1450亿美元，市场发展潜力较大。

全球首款

7月4日，据光明日报报道，北京大学集成电路学院教授杨玉超团队联合中国科学院上海微系统与信息技术研究所研究员宋志棠团队，成功研制出全球首款基于相变忆阻器的神经动力学系统芯片，首次将这类复杂运算的单步时延压缩至2.12毫秒，在脑皮层重建等任务中较目前先进图形处理器（GPU）提速达50至478倍，一举突破了制约神经动力学长达半个世纪的实时计算瓶颈。相关成果3日凌晨发表于《科学》。

杨玉超表示，要让机器像大脑那样实时建模和理解物理世界，需要一种将神经网络与微分方程相结合的“神经动力学系统”。它能在不完整、带噪声的数据中重建出平滑精确的三维脑结构，应用潜力巨大。

然而，传统计算架构存在一个核心瓶颈：存储与计算分离，求解过程中海量的中间变量在内存和处理器之间反复奔波，如同一个庞大的数据工厂，大量时间被浪费在搬运路途中，不仅延迟巨大，功耗也居高不下。

面对这一难题，研究团队从忆阻器本身的物理特性里找到了破局答案。他们利用相变存储器独特的“电导漂移”现象——在一定时间窗口内，其电导变化是可预测、可精准调控的。

基于此，团队提出“可控存内计算”新范式，将动力学系统求解中最耗时的自适应步长搜索，直接编码为器件的物理电导演化过程，在存储单元内部原地完成计算。通俗地讲，原本需要复杂数字电路反复执行的运算、缓存访问、数据搬运等工作，现在交给了器件自身的物理规律去“跑”。

更值得关注的是，团队还将神经网络权重映射到相变存储器的多级电导态上，在同一个阵列内同步完成矩阵乘加运算。

两大核心计算任务由此被统一集成在总面积仅0.28平方毫米的存算阵列中。这颗采用40纳米工艺的芯片运行频率为50兆赫兹，单步积分仅需9级流水线，最终实现2.12毫秒的单次迭代时延，首次将神经动力学硬件推入毫秒时代。

“性能表现令人振奋。”杨玉超表示，在同等运算下，该芯片较当前最先进的专用加速器速度提升3.82倍至36.27倍，功耗降低11.75至24.73倍；在脑皮层表面高保真重建任务中，甚至比NVIDIA A100 GPU提速高达478.18倍。重建出的脑皮层网格平滑、拓扑一致，能精准刻画复杂的褶皱结构，并有效抑制传统方法中的伪影和自相交缺陷。

杨玉超说，这一突破为脑机接口和脑疾病诊疗开启了全新想象空间。未来，个体化、动态化的脑数字孪生成为可能，术中神经导航、阿尔茨海默症早筛及个性化干预等，将获得可实时运行的硬件底座。

脑机接口赛道加速爆发

脑机接口是连接人脑与外部设备的新型人机交互技术，也是当前全球科技竞争的前沿领域。这一技术可以构建大脑与机器的直接信息通路，实现神经信号的读取、解码与写入，推动碳基生物智慧与硅基人工智能深度融合。

今年以来，脑机接口产业迎来了前所未有的热度。3月，“脑机接口”首次写入政府工作报告，与未来能源、量子科技、具身智能、6G并列，被明确为培育发展的未来产业之一。

2025年北京、上海、山东已经出台了脑机接口专项政策。今年以来，广东、江苏、天津也相继推出了脑机接口相关专项政策。

中国信息通信研究院预测，2025年我国脑机接口市场规模约为30亿元，2030年预计将达到120亿元。

从脑机产业链价值分布来看，上游的电极、芯片等核心器件构筑较高的技术壁垒，是产业的价值高地；中游的系统集成与解决方案是实现技术向产品转化的关键中枢；下游多元化应用是产业增长的核心引擎，随着下游应用环节逐步成熟，市场有望从百亿级向千亿级发展。

下游应用方面，医疗健康领域增长最快，据麦肯锡测算，全球脑机接口在医疗应用领域的2030年市场规模有望达到400亿美元，到2040年突破1450亿美元，市场发展潜力较大。

## ## ## ## 责编：罗晓霞

排版：刘珺宇

校对：吕久彪

## ## ## 百万用户都在看

上调20%！芯片巨头，突传重大调整！AI芯片大单，密集落地 4分钟，封死涨停！商业航天，集体爆发！重磅催化临近？凌晨，利空突袭！“AI搅局者”，突然崩跌！发生了什么？刚刚，利好传来！芯片巨头，飙涨！日韩股市大逆转紧急预警！三大突发变数，集中袭来！韩国股市，出啥事儿了？

违法和不良信息举报电话：0755-83514034

邮箱：bwb@stcn.com

!Image 5: 图片

Key Quotes

> This breakthrough opens up a whole new realm of imagination for brain-computer interfaces and the diagnosis and treatment of brain diseases.

> Tasks that originally required complex digital circuits to repeatedly execute operations, cache accesses, and data transfers are now handed over to the physical laws of the device itself to 'run'.

全球首款忆阻器神经动力学芯片问世，脑机接口赛道加速爆发