|
|
楼主 |
发表于 2025-5-23 08:38:36
|
显示全部楼层
Coming back from spending a week in Taiwan speaking to partners in the $NVDA supply chain, incrementally bullish on the GB200 / GB300 ramp this year.
The demand we’re seeing from AI labs / CSPs continues to accelerate sharply after the release of reasoning models. (See Microsoft / AWS commentary / OAI commentary)
Nvidia’s ability to ramp production of its highly complicated GB NVL72 systems remained an open question, as the system is completely unprecedented, with each Blackwell GPU needing to see the 71 others simultaneously through the NVLink Switch and Backplane. An order of magnitude harder than the previous form factor with 8 GPUs talking to each other in 1 HGX box.
Taiwan engineers are working incredibly hard, 24/7 (some a bit depressed), which explains the pessimistic read from Taiwan supply chain checks thus far. But the hard work has yielded shortening of cycle times for GB200 which will increase finished product shipment rates.
A key point Jensen and the ODMs stressed was that the CSPs were leaning on the ODMs to take on more of the traditional testing and bring up that they did for Hopper. So actually when a GB200 is shipped to the end customer, it’s as close to plug and play as shipping a server ever was.
Hon Hai and Wistron came off most bullish about their production capability, writing their own tests and automating production. Also importantly, expanding geographically to produce these supercomputers in the US.
Expecting rack throughput to grow by over 100% m/m for this quarter - and when these GB200s are used for training and inference, will usher in another era of AI models altogether.
Models and software are like water, and they adapt / expand to fit the new vessel that is deployed.
GB300 is expected to ship in Q3 as well and will lean on many of the learnings of GB200 shipments and will be a much faster ramp. Seems like quite a battle between bearish investors in Asia / somewhat bullish investors in the US - curious to see how things unfold!
Not investment advice. All opinions my own and do not reflect those of Altimeter.
由
翻译自 英语
在台湾待了一周，与$NVDA供应链的合作伙伴交谈后，我对今年 GB200 / GB300 的增长越来越看好。
推理模型发布后，我们看到来自人工智能实验室/云服务提供商 (CSP) 的需求持续急剧增长。（参见微软/AWS 评论/OAI 评论）
Nvidia 能否提高其高度复杂的 GB NVL72 系统的产量仍是一个悬而未决的问题，因为该系统是前所未有的，每个 Blackwell GPU 都需要通过 NVLink 交换机和背板同时访问其他 71 个 GPU。这比之前在一个 HGX 盒子里安装 8 个 GPU 相互通信的方案要难得多。
台湾工程师们正以惊人的速度、全天候（甚至有些沮丧）地工作，这也解释了迄今为止台湾供应链检查结果的悲观程度。但这些辛勤工作已经缩短了GB200的周期，这将提高成品的出货率。
Jensen 和 ODM 强调的一个关键点是，CSP 更倾向于让 ODM 承担更多传统的测试和升级工作，就像他们为 Hopper 所做的那样。因此，当 GB200 交付给最终客户时，它实际上就像交付服务器一样，几乎可以做到即插即用。
鸿海和纬创对其生产能力最为看好，他们自主编写测试并实现了生产自动化。同样重要的是，他们正在拓展地域范围，在美国生产这些超级计算机。
预计本季度机架吞吐量将环比增长 100% 以上 - 当这些 GB200 用于训练和推理时，将开启 AI 模型的另一个时代。
模型和软件就像水一样，它们会适应/扩展以适应部署的新容器。
GB300 预计也将于第三季度发货，并将借鉴 GB200 发货过程中积累的诸多经验，并将以更快的速度增长。亚洲看跌投资者和美国略微看涨的投资者之间似乎存在着一场激烈的较量——我很期待事态发展！
并非投资建议。所有观点均为我个人观点，不代表 Altimeter 的观点。 |
|
IP卡
狗仔卡
显身卡