战狼2

AI交换芯片走到聚光灯下：英伟达发布Spectrum-6 或推动Token生成能力倍增_我的网站

A |     NORRISTOWN, Pa. -- A man has pleaded guilty to third-degree murder in the death of a Philadelphia steakhouse chef prosecutors say he intentionally killed during a drug-fueled, hit-and-run spree that caused injuries and damage in two counties.Kareem Welton, 43, also pleaded guilty Friday in Montgomery County Court to attempted murder, three counts of aggravated assault and robbery of a motor vehicle in the violent predawn 2021 joyride that started in Philadelphia's Center City and ended 30 miles (48 kilometers) away in neighboring Montgomery County.Prosecutors dropped a first-degree murder charge in exchange for the plea, which does not include any agreement on sentencing.Authorities alleged that Welton stole a car left running by a woman in July 2021 and then went to a family friend's home and stole a 5-year-old poodle, hit two people on a motorcycle and then struck and killed 31-year-old Adriana Moreno-Sanchez as she walked home from her restaurant shift. Authorities said in an affidavit of probable cause that the car hopped the curb and struck Moreno-Sanchez, then made a U-turn and went back onto the sidewalk to hit her a second time. Authorities said Welton briefly stopped at that time to steal a jacket from the injured victim, who died at a hospital.Authorities said Welton drove off, striking more cars and people, before abandoning the car. He later allegedly tried unsuccessfully to pull another woman from a vehicle before stealing another car blocks away and heading out of the city, hitting a bicyclist on the way. Eventually, he was spotted in Collegeville, and during the ensuing chase authorities allege he deliberately swerved into oncoming traffic to hit and critically injure a jogger. He was arrested after pulling into a Trappe gas station and trying to take a bicycle from a man standing nearby, authorities alleged.A toxicology screen later indicated that Welton had marijuana and PCP, a powerful hallucinogenic, in his system at the time of the assaults, authorities said.Both the 3rd-degree murder and attempted murder charges carry maximum sentences of 20 to 40 years in prison. Assistant District Attorney Roderick McCord Fancher III told The (Pottstown) Mercury that he would seek significant prison time including consecutive terms, a sentence “that accounts for each and every victim of this rampage.”"He brought people’s worst fears of random acts of violence into real life through his actions. Completely innocent people going about their day were seriously affected by him,” Fancher said. Montgomery County authorities, using a state law, prosecuted all charges, including the fatal hit-and-run that occurred in Philadelphia.A message seeking comment was sent Saturday to the assistant public defender representing Welton.。    《科创板日报》7月22日讯 近日，英伟达正式发布新一代Spectrum-6以太网交换芯片，用于支持基于Vera Rubin平台的超大规模AI基础设施。         英伟达表示，Spectrum-6交换芯片将作为下一代Spectrum-X Ethernet交换机平台核心组件，通过高带宽、智能流量管理和故障恢复能力，提升AI工厂整体计算效率。包括CoreWeave 、微软、Nebius、SpaceXAI和特斯拉在内的AI基础设施建设者，将成为首批引入Spectrum-6来加速其AI工厂建设的公司。         产品性能方面，Spectrum-6单交换容量达102.4Tbps，是上一代产品的两倍，旨在连接数十万级GPU和CPU，为大规模AI训练、智能体应用及推理任务提供网络支持。基于此，Spectrum-X以太网交换机能够实现比普通以太网高1.6倍的AI网络性能，并且在超过10万个GPU的部署中保持高达95%的网络效率。         此外，Spectrum-6交换芯片同时支持可插拔式光模块和共封装光学（CPO），并支持液冷散热，可为整个AI工厂提供全面的端到端散热方案，同时提高网络电源效率。         截至目前，基于Spectrum-6的交换机已经进入全面生产阶段，且在全球千兆级AI工厂中普及。         中国银河指出，Spectrum-X Ethernet Photonics是全球首款200G/lane大规模量产的CPO交换机。Vera Rubin量产Spectrum-X Ethernet Photonics进一步验证光互连在下一代AI数据中心中的价值。硅光子技术渗透率的提升也将同步带动上游光芯片需求的高速增长。         ▌交换机成Token生成倍增器          交换机是一种用于电光信号转发的网络设备，而在英伟达看来，AI性能本质上是网络问题。         其指出，全球最先进的人工智能工厂汇集了数十万个GPU和CPU，用于训练前沿模型、驱动智能体人工智能，并以前所未有的规模生成智能。在这个层面上，网络成为推动Token生成的关键计算能力倍增器。

B |          对交换机领域的重视，也让这家科技巨头收获了客户与收入增长。         作为Spectrum-6交换芯片的首批用户，CoreWeave表示：“将英伟达 Spectrum-6和液冷式Spectrum-X以太网基础设施引入AI工厂，有助于我们为客户提供训练前沿模型和更快部署推理所需的带宽、弹性和效率。”          据IDC最新数据，英伟达在2026财年第一季度首次成为全球数据中心以太网交换机市场的营收冠军。在过去，该领域由博通、思科等网络巨头主导。         开源证券表示，从交换芯片和GPU配比来看，传统架构交换机与GPU配比约为3:64，而英伟达NVL72包含72颗GPU和9个交换托盘，GBNVL72柜内交换托盘各配备2颗交换芯片，交换芯片与GPU配比达1:4，最新VRNVL72柜内交换托盘配备交换芯片数翻倍，交换芯片与GPU配比进一步提升至1:2，随着Scaleout和Scaleup双线演进，交换网络在计算集群中所占权重或逐步提升。

C |          该机构强调，交换网络正算力配套设施配角，跃升为集群核心主角。随着AI模型参数扩张，算力需求已从单纯GPU堆叠，转向全维度系统架构重构。受单芯片物理功耗密度、互连带宽及内存容量瓶颈制约，算力增长边际效益持续递减，交换网络成为制约算力的天花板，当前系统级协同架构（如高带宽域互联）是突破单芯片性能上限的主要技术路径。         （科创板日报 张真）。

Current article:http://ovul.diuhuairandaikaomangshangnen.buzz/x3rk0/20260826/84556.html

Published on:00:00:00