A | “Following the Fourth of July shooting in Highland Park, IL, Chair @RepMaloney sent letters to three major gun manufacturers requesting their appearance at an @OversightDems hearing on July 20, 2022,” the committee wrote on social media.,The letters were addressed to Marty Daniel, the CEO of Daniel Defense; Mark Smith, the president and CEO of Smith & Wesson Brands; and Christopher Killoy, the president and CEO of Sturm, Ruger & Co., The Washington Post reported.,This hearing is the second held by the committee to investigate the gun industry. Witnesses of the Uvalde and Buffalo mass shootings in May shared their experiences at the June 8 hearing. When the first hearing was announced in early June, the committee noted there had already been more than 200 mass shootings in the United States this year.,Committee Chairwoman Carolyn Maloney is requesting the presence of the CEOs at the second hearing because of financial information given to the committee, according to The Washington Post, which acquired a copy of the letter sent to Killoy. She had previously requested five gun manufacturers to provide their gross revenue and profit for sales of AR-15-style semiautomatic rifles; annual advertising and marketing expenses for these guns; annual spending on federal and state lobbying, and funding for the National Rifle Association.,Maloney said the committee’s concerns that gun manufacturers were not taking sufficient precautions to limit the harm caused by their rifles had only grown after receiving the information. In a statement cited by The Washington Post, she said the CEOs needed to explain why they were selling rifles designed for the battlefield to civilians.。 8月26日消息，三星宣布，全球首次研发出面向移动端及端侧AI设备的下一代内存——LPDDR-PIM，该技术将运算功能集成于移动内存中，可有效提升AI处理速度并降低功耗。
在美国斯坦福大学举行的半导体学术会议Hot Chips 2026上，三星相关负责人首度公开展示了LPDDR5X-PIM的结构、规格及验证结果——该产品本月初刚在FMS 2026内存会议上斩获下一代内存奖项，但详细参数直到此次会议才正式披露。
B | PIM（内存内运算）技术通过在内存中嵌入运算单元，使部分计算可直接在内存内部完成，无需像传统架构那样将数据频繁搬运至CPU或GPU再传回内存。
C | 对于AI运行而言，数据搬运是主要性能瓶颈之一，应用PIM后，简单运算就近处理，从而在提升速度的同时显著降低功耗。
在测试环节，三星以参数规模达80亿的Meta开源AI模型Llama 3.1为基准，结果显示LPDDR5X-PIM的处理时间为5.4秒，较普通LPDDR5X的12.3秒快了约2.3倍；每秒令牌处理量达81.3 TPS，为现有产品27 TPS的三倍。 事实上，三星早在2021年便公布了面向高带宽内存的HBM-PIM技术并完成开发验证。此次将PIM落地于智能手机等移动设备广泛采用的LPDDR低功耗内存，正是顺应大语言模型向小型化、高效化演进的新趋势。
D | 相关负责人指出，虽然HBM是目前AI内存的标准方案，但并非所有场景都非它不可，LPDDR-PIM兼具高带宽与低功耗，为端侧AI开辟了新的可能。
随着生成式AI直接在智能手机、PC等终端设备上运行的端侧AI趋势不断扩展，LPDDR-PIM的应用前景持续看涨——电力与空间受限的移动设备，亟需在提升AI算力的同时严控功耗。 在兼容性方面，该产品沿用与通用LPDDR相同的阵列设计，无需大幅改动系统即可部署。目前，三星正提供软件开发工具包和模拟器等资源，积极推动生态建设。 继LPDDR5X-PIM之后，三星已启动LPDDR6-PIM的标准化工作，AI内存竞争正从以HBM为中心的数据中心，延伸至智能手机和PC等端侧领域，内存也正从单纯的数据存储设备，向直接参与运算的关键组件演进。
Current article:http://rnd3m3y.cuozubeishenjingli.cfd/news/20260826_1901565.html
Published on:19:03:22
