您的位置：首页 > 新闻中心 > 社会 > 正文

射雕英雄传83版

Nonprofit seeks insight from devices left behind by veterans who died by suicide_我的网站

半泽直树

一 | 8月24日,小米在北京召开玄戒芯片技术沟通会，正式发布玄戒O3、O100和D100三款自研芯片，分别面向旗舰手机SoC、端侧AI加速和智能驾驶等场景，“三芯齐发”共同拼出"从口袋到座舱、从客厅到工厂"的人车家全生态AI算力版图。 玄戒O3：当代AI旗舰SoC，安兔兔跑分超520万 O3是小米自主研发设计的第二款高端旗舰SoC，延续3nm制程，晶体管从O1的190亿增至240亿。 The cell phones, tablets, and laptops left behind by veterans who died by suicide may offer insight that could help prevent future military suicides, according to a nonprofit dedicated to analyzing the data from the devices.Stop Soldier Suicide collects the technology as part of its initiative, "The Black Box Project," which is named after the black boxes found on planes that help determine the cause of aviation incidents."If we could gather novel data from those that have died by suicide, we might be able to recreate the last year of life," Stop Soldier Suicide CEO Chris Ford told ABC News.Veterans are 57% more at risk for suicide than non-military adults and around 6,000 veterans die by suicide each year, according to the Department of Veterans Affairs."I think we can prevent needless suffering and a ton of loss of life," Ford said.That kind of loss is something David McDaniel, whose son Connor died by suicide at the age of 26, knows all too well.Connor McDaniel joined the Army after high school, and, according to David, he appeared upbeat during his deployment to Afghanistan, where he was assigned to protect local villages from the Taliban."Some of his Army buddies told me that he was making them laugh while they were being shelled in Afghanistan," David McDaniel told ABC News.But it wasn't long after Connor came home that David McDaniel says he sensed a shift in his son's behavior to include "more anger, a lot of frustration, a lot of gloominess."On Sept. 1, 2021, David says his son entered a wooded area with a firearm."I start screaming his name as I'm walking up the field. And just then the police, the sheriff's department shows up and they tell me to come back down," McDaniel said.He followed their direction, but on the car ride home, an officer called him and said Connor was dead after he fired upon sheriff's deputies and the deputies returned fire and killed him.Connor's death was determined to be a suicide. The district attorney's office concluded that the officers' actions were justified.While grappling with his son's death, David McDaniel was immediately drawn to the Black Box Project, which uses forensic software to unlock and comb through service members' devices provided by their loved ones to review their activity in the weeks leading up to their suicides.McDaniel works in the tech field and says if the project can find repeated patterns in online searches, they may be able to help others to prevent suicides.For instance, McDaniel says his son searched for SGLI, Servicemembers' Group Life Insurance, before his death. If a veteran were to make a similar search and look up "suicide" in the future, for example, McDaniel suggests it could trigger an ad for Stop Soldier Suicide to pop up and provide resources in the search results.Ford says the nonprofit uses machine learning, a subfield of artificial intelligence, to analyze texts and notes in order to determine sentiment or emotional tone, such as anger or desperation. He says they can also track phone usage of the deceased veterans to monitor their past sleep cycles and trace their past geo-locations."Machines can go through computations a lot faster than humans, and they can find signals that we often miss," Ford said.Since launching the Black Box Project in April 2020, Stop Soldier Suicide has received over 100 devices, Ford said. So far, he says they're seeing three patterns before most of these suicides: more anger, different sleep patterns, and social isolation.Ford's team includes Ph.D.s in clinical psychology and suicide researchers. A scientific advisory committee helps steer where and what to look for on the devices.The findings are still preliminary, according to the nonprofit, but one goal is to help the organization decide where, when, and how to target their own mental health resources, like adding more hotline staff at certain hours and ensuring calls are picked up overnight.When asked about privacy concerns of collecting data, Ford said it's something they think about daily."We don't resell the data. We don't push it out to others. We use it for our own research purposes so that we can best understand risk and then give those devices back to those families intact," Ford said.Current privacy laws do not carry on past death, meaning the next of kin typically gains new authority.Petra Oxford-Jackson, whose veteran husband Anthony died by suicide, says she gave his phone to the organization with the possibility it could save even one life."I truly believe he would have been okay with it, and it's a hard thing for a survivor to give up something so personal as his cell phone," Oxford-Jackson told ABC News.The Black Box Project received $3 million in funding from the Department of Veterans Affairs earlier this year.If you are struggling with thoughts of suicide or worried about a friend or loved one, call or text the Suicide & Crisis Lifeline at 988 for free, confidential emotional support 24 hours a day, seven days a week.。小米实验室给出的安兔兔综合跑分为5228014，是行业内首个突破500万分的移动旗舰平台。此外，新芯片在架构层面也几乎全面重做，其中CPU部分采用6超大核+4大核的十核“全大核”架构，峰值主频4.35GHz。 GeekBench 6单核3945分，多核15221分，相较上一代分别提升31%、60%，涨幅十分显著；同时，日常使用场景功耗则进一步降低25%，兼顾性能和能效表现。 图形方面，O3首发新一代G2-Ultra NX图形处理器，延续16核超大规模配置。

二 | 传统图形性能较O1提升85%，光线追踪性能提升182%，相同性能下功耗至多降低64%；其中，Aztec帧率213帧，较上代提升94%，堪称跨代式的提升。影像上，玄戒O3的提升也是巨大的，第五代ISP单摄支持4.32亿像素镜头，内部图像信号处理位宽达24-bit，AI RAW域降噪夜景视频规格提升至4K/60FPS；DPU将屏幕分区色调映射能力拓展至更多场景，提升暗光可读性、增强HDR片源显示效果；VPU采用编解码分离架构，视频剪辑导出速度提升20%，并在安卓领域首发H.266硬件解码。另外，安全模块全栈重构，获国密与CCRC EAL5+安全认证；基带通信性能看齐主流旗舰平台集成方案，5G场景综合功耗相比前代降低20%以上。内存与缓存这次也有不小的升级。O3在片上集成了合计60MB缓存，其中16MB SLC补上了前代短板；同时首发支持LPDDR6，内存带宽113.8GB/s，比上代高出48%。

三 | 缓存和内存带宽升级之后，玄戒O3也对芯片内部的数据传输进行了调整。

四 | 它采用统一融合总线架构，减少不同模块之间的协议转换，同时配合顶层金属层数据通路和数据预取技术，数据显示，统一传输协议可降低75%的协议转换开销，内存访问时延低至82ns，达到业内领先水准。而在调度上，小米也强调，日常卡顿里约有三成并不是算力不够，而是任务在排队等资源。玄戒O3为此加入了硬件调度控制单元，配合全栈QoS机制，让关键任务在每一个环节都能优先通过；再结合对负载、温度、电压的实时监测和场景预测，全局资源反馈最快1ms就能完成，高频使用场景的能效收益普遍在20%以上。在大家重点关注的AI上，玄戒O3在CPU、GPU等主要模块中加入了AI计算单元，用于处理翻译、音视频理解、图形处理等轻量AI任务，减少跨模块调用带来的数据传输和额外功耗。当然，大模型部分依旧由NPU负责，玄戒O3采用4核NPU，Tensor算力达到200TOPS，并加入SIMT架构Vector单元，提供3.13TFLOPS向量算力。针对端侧大模型，玄戒O3还与Xiaomi MiMo进行了联合优化，通过5值量化和硬件霍夫曼无损压缩降低模型数据量，根据发布会现场公布的数据显示，模型推理速度相比主流旗舰SoC提升45%。 玄戒O100：1.22TB/s高带宽，夯实Agent时代硬件基础 除了玄戒O3之外，这次玄戒芯片技术沟通会还发布了两款新芯片：玄戒O100和玄戒D100。其中玄戒O100是一颗6nm端侧AI加速芯片，可与玄戒O3配合运行端侧大模型，在弱网或无网环境下完成本地AI计算。 O100采用3D Wafer On Wafer立体堆叠，通过Hybrid Bonding技术将两层DRAM晶圆与一层NPU计算晶圆连接在一起，芯片内部共有258万个键合节点，键合间距为1.4μm。顶层DRAM与计算晶圆之间布置了28672条连接通路，相比传统外置内存方案，可以提供更高的互联密度和更短的数据传输路径。按照小米公布的数据，玄戒O100内存带宽达到1.22TB/s，约为主流手机内存带宽的16倍，端侧大模型推理速度最高可达330Token/s。芯片内部还配置14核大模型专用NPU，并采用XRING HB-Matrix高带宽矩阵总线。

五 | 在工艺方面，O100没有采用传统MicroBump方案，Hybrid Bonding的TSV直径做到0.7μm，三层芯片之间还采用F2F金属层直连结构。高温键合过程中容易出现晶圆翘曲和碎裂，小米表示，团队为此用了约半年时间迭代芯片版图，并围绕F2F金属层直连架构进行了多项专利布局。 玄戒D100：国内首个3nm智驾芯片，端侧跑200B大模型 玄戒D100采用3nm工艺，定位智驾高算力AI芯片，内部配置20核高性能CPU和16核NPU，是国内首款3nm智驾芯片。除了运行智驾算法，D100还支持更大规模的本地AI计算。

六 | 单颗芯片最高支持160GB内存，可部署超过200B参数规模的大模型；多颗D100还可以通过玄戒高速互联总线进行联合计算。 写在最后： 按照目前公布的计划，玄戒O3将于今年9月随Xiaomi 18 Fold首发上市；玄戒O100和D100均已完成研发，预计明年正式商用。一年前，玄戒O1的发布证明了小米已经具备自研旗舰SoC的能力；如今O3、O100、D100三款芯片同时亮相，玄戒的应用范围也从手机扩展到了端侧AI和智能驾驶，开始覆盖小米“人车家”全生态的算力需求。自研SoC从来不是一条轻松的路线，尤其到了旗舰级别，研发、量产、调校和后续迭代都需要长期投入。