hold

A hug across 27 years：American teacher and Chinese desert fighter reunite and embrace with tears in China's Inner Mongolia_我的网站

A |     8月24日消息，今天下午，小米除了发布玄戒O3之外，还带来了玄戒O100，一款大模型专用的AI加速芯片，更是全球首款6nm 3D晶圆级堆叠的AI加速芯片。    玄戒O100通过先进的3D堆叠工艺，实现了1.22TB/s的超高带宽，这一速度是目前主流旗舰手机的16倍之多，可以让移动端设备的AI性能实现质的突破。玄戒O100与玄戒O3是AI终端的“最佳CP”，最高可以实现本地AI推理速度高达330Tokens/s。    

    Ronald Sakolsky (center) reunites with sand fighter Yin Yuzhen on August 23, 2026 in Ordos, North China's Inner Mongolia Autonomous Region. Yin received his $5,000 donation and turn them into 50,000 full-grown trees. Photo: Lin Xiaoyi/GT
    
A long-awaited reunion that captured nationwide public attention unfolded on Sunday evening: American teacher Ronald Sakolsky and Yin Yuzhen, a model worker renowned for planting trees and combating desertification, reunited and embraced tearfully with overwhelming emotion at an airport in Ordos, Inner Mongolia, across 27 years and the separation of the ocean. 
In 1999, Sakolsky raised $5,000 to fund Yin's tree-planting initiative. The young saplings bought with the donation have since matured into more than 50,000 full-grown trees in the once-barren Mu Us Desert in Inner Mongolia. 
Yin waited at the arrival gate very early, the Global Times reporter observed, as she murmured English sentence she particularly learned for this occasion "Welcome to China my brother." She wore an outfit with green shade, as "green is the color of the Earth." 
According to Yin, she and her family has prepared for the reunion steamed buns, meaning "upward," and  vegetables and fruits picked from the land transformed from desert.
Yin's eyes were fixed on  the escalator, and wiped out tears multiple times during the waiting time as travelers come in groups. 
Sakolsky said in earlier interview with the Global Times that he looked forward to this reunion and cannot hold his tears when he thought of the moment. He was about to give Yin a hug when they meet. 
When Sakolsky finally appeared in front of Yin, both people's eyes welled up with tears. Stepping off the elevator, Sakolsky could barely contain his excitement, muttering "Oh my God, oh my God." They hurried toward each other and locked in a tight embrace at last - the long‑awaited hug 27 years in the making, the Global Times saw. 
"This is a dream come true, from a lady who is the hero of the story, not me," Sakolsky told the Global Times at the airport, barely holding his emotion. "She made a promise 27 years ago, and she kept her promise. I never thought I would see her again. It is a miracle." 
"Brother" is one of the words Yin specially learned to greet Sakolsky and the American friend told media that "we are brother and sister now. 27 years ago, we were just friends and acquaintances."   
So much has happened since then, but the important thing is many people have seen the story. And now they can also help to make China beautiful, Sakolsky noted. 
"When I hugged her the first time, she stood like this, she was very scared. Chinese women do not hug men especially foreigners in public. But this time is exception, we are brother and sister now," he said.  
。     传统计算架构下，计算单元与存储单元物理分离，数据必须在两者之间反复搬运。但内存数据通路受限于物理引脚数量，遇到了增长瓶颈。    当AI算力以指数级增长时，内存访问速度与算力之间已经严重脱节，成为性能的关键阻碍，业界称之为“内存墙”。玄戒O100，则是小米“推倒内存墙”的破局之法。     区别于以往晶圆→切割为裸片（Die）→逐片封装互联的方式，玄戒O100采用先进的Wafer on Wafer封装技术，也就是将多片完整的晶圆直接高精度对准与键合，继而再切割成独立的3D芯片。    这彻底摆脱了传统引脚数量对带宽的物理限制，将数据通路从“平面走线”变为“垂直穿透”，实现真正的近存计算（Near-Memory Computing）。         了解大体逻辑后，我们看看玄戒O100到底如何做的。首先将两层AI专用高速DRAM晶圆与一层高性能NPU晶圆垂直堆叠，然后在垂直方向“钻隧道”让数据可以高速流转。

B |     所以，只要隧道越多，也就可以实现越高的带宽。为此，玄戒O100采用了先进的Hybrid Bonding混合键合工艺，摒弃了传统微凸点（Micro Bump）结构，直接将物理通路的间隔缩短至1.4μm，让数据通路密度大幅提升。

C |               为了进一步缩短DRAM与NPU计算层的物理距离，玄戒O100还采用了 Face to Face（简称F2F）金属层直连结构。    简单来说，每一层晶圆内部，其实底部是逻辑晶体管，上层有密布多层的金属走线层。F2F，就是让DRAM与NPU的金属走线“面面相对”，如此可以实现更短的物理互联距离，实现更快的信号传输。    当然，这说起来简单，在6nm如此高工艺节点下，要解决散热、电磁干扰等诸多难题，玄戒团队逐一攻坚，实现最终量产，并申请了15项专利，其中已获批4项发明专利。         与此同时，玄戒O100配备了14颗高算力NPU，通过玄戒高带宽矩阵总线，让数据可以在NPU与内存、与其他NPU核心之间等高速流转。

D |     创新的矩阵总线支持拓扑动态切换，Prefill阶段采用环形拓扑，数据逐核流转、顺序同步；Decode阶段切换为广播拓扑，所有结果统一汇聚到指定核心。总线控制上的灵活拓扑，让片上计算资源得以充分应用。

E |          最终，玄戒O100的带宽达到了惊人的1.22TB/s。大家对这个数字可能没什么概念，以目前主流旗舰手机LPDDR5X内存作为对比，它的带宽76.8GB/s，而玄戒O100是其16倍之多，彻底释放端侧AI性能。         为了验证玄戒O100的高带宽表现，小米工程师还做了一款原型机，搭载玄戒O3和O100，双芯协同计算。

F | 小米还引入了风冷主动散热，充分发挥芯片的全部性能。配合Xiaomi MiMo端侧模型，实现了超高速端侧推理性能。                   

。

Current article:http://1md.rangzhuanhoujuedu.cyou/k98jng/20260826/4072.html

Published on:01:02:51