“榨”出硅的极限：怎么让GPU不“闲着”？｜与SGLang、RadixArk深聊AI Infra技术与千亿美元市场

OpenAI、谷歌、Deepseek...各大AI前沿模型最近开始了一场"造芯大战"，市场依旧缺卡缺算力，因而如何让手里的GPU发挥出更大价值，正在成为硅谷AI Infra（基础设施）背后千亿级的热门赛道。当推理成为AI的主战场，GPU看似满载，却仍把大量时间耗在重复计算、缓存搬运和任务等待上。AI Infra深处，一场围绕调度与系统优化的效率革命已经开始。为什么最昂贵的GPU仍会“又忙又闲”？AI Infra还能榨出多少藏在“硅”里的性能？当算力增长不再只靠堆卡，AI竞争的尺度会被怎样改写？这期视频，我们与推理引擎开源社区SGLang孵化出的RadixArk团队一起，深入探讨这场算力效率变革。采访嘉宾：朱邦华，RadixArk联合创始人、英伟达前首席研究科学家陈震林Richard，RadixArk Member of Technical Staff Ethan Xu，前微软能源战略经理、突破能源科研总监【欢迎加入硅谷101 Discord讨论群】 👉https://discord.gg/PACWRvyfCm 时间轴： 00:00 榨出“硅”的极限：重新定义一块GPU 01:43 昂贵的GPU，大多数时候是很“闲”的 04:21 AI Infra的四层架构 08:45 优化“计算”：SGLang、Radix Tree与K/V Cache的极致复用 14:43 优化“等待”：PD分离 16:54 优化“算力”：多释放一倍算力的两条路径 21:03 资源协同：适配RL后训练的Miles 25:50 RadixArk的更大野心：“以后没有Frontier Lab这个词” 【关于硅谷101】我们是由海内外一线媒体记者/主持人创办的栏目，深度解析硅谷创新趋势，以轻松的风格分享科技行业的最新动态。我们采过顶级科技大佬，积累了数万小时的媒体经验，做过调查性报道，操盘过千万级传播量的知名深度稿，引发全国讨论和微博热搜；致力于将最专业的媒体素养和信息搜集能力转化为易传播的新媒体力量。旗下同名播客栏目：https://www.youtube.com/@valley101podcast 关注我们，从这里驶向未来。联系我们：video@sv101.net 【往期精彩视频】 https://www.youtube.com/watch?v=Hq4jFGd4YxE https://www.youtube.com/watch?v=jXZGeF6sJOc https://www.youtube.com/watch?v=nilNLfW7izg&t=2252s https://www.youtube.com/watch?v=i0t3RyLILnw https://www.youtube.com/watch?v=erIEZkMd9_w https://www.youtube.com/watch?v=9mpazu2K3qU https://www.youtube.com/watch?v=jnYWv_0MzBw https://www.youtube.com/watch?v=ULi-TCF2cQE&t=5s https://www.youtube.com/watch?v=nWc1tpQnhqY https://www.youtube.com/watch?v=y2bZdXZBeBc https://www.youtube.com/watch?v=hIdvhl8PUwc https://www.youtube.com/watch?v=x2meJPOn9ws https://www.youtube.com/watch?v=HqsTB0avrh8 https://www.youtube.com/watch?v=F7oNfszczVc https://www.youtube.com/watch?v=DZIhQeFEdXI https://www.youtube.com/watch?v=1Rh7Zwd7Vx0 https://www.youtube.com/watch?v=cqW_VWYbIcU https://www.youtube.com/watch?v=9RMvKWHgyF8 https://www.youtube.com/watch?v=qs0ovkaacls https://www.youtube.com/watch?v=g5hw1HZfwTc https://www.youtube.com/watch?v=ppRvzPXGpEw https://www.youtube.com/watch?v=HLJdElh2tFc https://www.youtube.com/watch?v=jn_H3vuUC6E https://www.youtube.com/watch?v=CIDVbaXWp64 https://www.youtube.com/watch?v=7D_8zimKtB4 https://www.youtube.com/watch?v=tzEWYNQmnmc https://www.youtube.com/watch?v=kjqOsLdvSFg https://www.youtube.com/watch?v=kjqOsLdvSFg https://www.youtube.com/watch?v=HjopX519nXg https://www.youtube.com/watch?v=pdcT2jwXP0s https://www.youtube.com/watch?v=uCvqXv_FJtI https://www.youtube.com/watch?v=SYuSZIIYOfI&t https://www.youtube.com/watch?v=Tzp0EF-sysI https://www.youtube.com/watch?v=CfxuqbB9dj4 https://www.youtube.com/watch?v=ZDXIUO7CRzo https://www.youtube.com/watch?v=18qSqzTdnbs https://www.youtube.com/watch?v=j13RPyHHCWk https://www.youtube.com/watch?v=VJgIENvy14k https://www.youtube.com/watch?v=thUc2wdtUTY https://www.youtube.com/watch?v=Hemlsyob1Ng https://www.youtube.com/watch?v=bO6xBjIx9JQ