火星情报局
小米玄戒O100推倒内存墙:实现真正的近存计算_我的网站

A | New Delhi, Oct 12 (UNI) Indian Army has launched a drive to introduce Electric Vehicles (EVs) into select units that will significantly reduce the carbon emissions and dependency on fossil fuels.
According to the plans, a detailed roadmap has been prepared to ensure that 25 per cent light vehicles, 38 per cent buses and 48 per cent motor cycles of the select Army units are changed to Electric Vehicles soon with adequate charging infrastructure, officials said.
In April 2022, the Indian Army had organised a demonstration of available EVs to Defence Minister Rajnath Singh where Electric Vehicle manufactures from Tata Motors, Perfect Metal Industries (PMI) and Revolt Motors showcased their EVs and briefed about the enhancement in technology and range of operation achieved during past few years.
The drive to introduce Electric Vehicles in the Indian Army gained further momentum after the demonstration.
Various factors unique to Indian Army’s employability, remote locations of employment and operational commitments were considered to arrive at a definite time bound roadmap.
Introduction of Electric Vehicles is considered as one of the effective steps to reduce Carbon Emission and reduce dependency on fossil fuel.
"India has emerged as one of the leaders to exceed global targets of reducing greenhouse gas emissions by 2030 and 2070. Many projects are being implemented to reduce the carbon emissions and dependency on fossil fuels," officials said.
In order to support a viable Electric Vehicles ecosystem in the Army, support infrastructure is also being established like EV charging points in the parking lots of offices/ residential complexes for on board charging.
These EV charging stations will have at least one fast charger and two to three slow chargers.
Besides, electric circuit cables, Transformers with adequate load bearing capability based on anticipated number of EVs per station are being ensured.
Solar Panel Driven charging stations are also planned in a phased manner to bring carbon footprint of these EVs to near zero.
Keeping in view the necessity and employability of EVs over various terrains, the Army will equip few units located in Peace Stations with EVs sequentially.
The Indian Army is also procuring EVs through the Capital Route.
As per plans, the existing deficiency of Buses will be fulfilled by procuring Electric Buses for select Peace Establishments for initial exploitation.
An Open Tender Enquiry for procurement of 60 Buses (Electric) along with 24 Fast Chargers will soon be floated, officials said.
The drive for Electric Vehicles will establish the right momentum for setting up the ecosystem for further induction of EVs in the Armed Forces, officials added.
Apart from these, Army has already started using EVs as part of Civil Hired Transport (CHT). Stations like Delhi Cantonment have already established charging stations to support EVs being hired or inducted subsequently.
At Delhi Cantt, a number of charging stations is also open to civilians.
UNI ASU SY ING。 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。
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。

B | 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”。

C | 玄戒O100,则是小米“推倒内存墙”的破局之法。

D | 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。

E | 所以,只要隧道越多,也就可以实现越高的带宽。

F | 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。 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。

G | 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。 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。 最终,玄戒O100的带宽达到了惊人的1.22TB/s。

H | 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。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://y1kx1u.xuebenwagangqingre.buzz/cyvq7/2d3b.html
Published on:02:04:38
