11-19 更新 实时
首页/ 算力/ 正文

NVIDIA 与 Lambda 披露首个量产共封装光学集群

算力 来源:ServeTheHome 2026-11-17 05:24 0 阅读 机器翻译
NVIDIA 与 Lambda 披露首个量产共封装光学集群

在 SC26 上,NVIDIA 与 Lambda 讨论了他们的首个量产级共封装光学(CPO)集群。业界为把 CPO 从实验室推向量产已经努力了数十年。业内很多人会告诉你这还需要几年时间,但 Lambda 已经在一个拥有超过 10,000 块 GPU 的集群中将其投入生产部署。

NVIDIA 与 Lambda 披露首个量产级共封装光学集群

我们在 NVIDIA 的 Hot Chips 2025 演讲中对其共封装光学做了详尽报道。NVIDIA 该封装的第一版应用于其 Quantum-X 交换产品,而 Spectrum-X 将是更新的第二代产品。

Hot Chips 2025 上 NVIDIA 采用硅光子的共封装光学

如上图所示,NVIDIA 在机箱正面使用由 18 个可插拔光源组成的阵列,来产生通往交换机封装光学模块的光。这些光源可插拔,意味着它们可以维护,而在光源被集成的方案中,这正是 CPO 应用于交换机的关键阻碍之一。在可插拔光学器件中,光源往往是容易失效的部件,因此让其可维护是一项重要的创新。同样,在交换机封装这一侧,让连接至少在一定程度上可维护,也有助于部署。

Hot Chips 2025 上 NVIDIA 采用硅光子的共封装光学

在下游,上图所示的 NVIDIA Quantum-X CPO 交换机拥有 144 个 MPO/MTP 端口,每个端口速率为 800Gbps。有意思的是,由于只有 18 个光源,而非 144 个 800Gbps 光模块(或 72 个双 800G 模块),可能失效的光源实际上比传统交换机更少,而且相比使用更大规模的独立可插拔模块阵列,它们也更容易散热。

该部署带来的最大好处之一是:通过把光学器件移到交换机封装上,而不是采用前面板可插拔模块,众多高速铜链路无需再从交换芯片出发、长距离地把信号驱动到前端笼架。因此,CPO 可以显著降低功耗,从而在相同的占地空间里容纳更多 GPU。毕竟,在功耗受限的环境中,最终能有更多电力直接用于 GPU,会直接影响到集群每瓦可用电力所能产生的收入。

Patrick,加拿大魁北克城,2026 年 11 月,身穿黄色 STH Tour 拍摄 T 恤

Lambda 和 NVIDIA 也许已经放出了关于该集群的短片,但我要明确一点:它是真实存在的,而且今天已经在运行,这让 NVIDIA 站在了部署 CPO 的最前沿。我认为它其实只是本周作为 SC26 的一部分才被宣布。说个有趣的插曲,几周前我在 OCP Summit 2026 上,听到另一家厂商在主题演讲中说,CPO 距离部署还有好几年。结果它其实已经部署了。我们不必再问 CPO 最终是否会部署、或者何时部署。现在的大问题是,它的采用曲线会是怎样的。

译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(NVIDIA and Lambda Disclose First Production Co-Packaged Optics Cluster… It is FINALLY real)

For SC26, NVIDIA and Lambda discussed their first production co-packaged optics (CPO) cluster. The industry has worked to move CPO from the lab to production for decades. Many in the industry will tell you it is a few years out, but Lambda has already deployed it in production in a cluster with over 10,000 GPUs.

NVIDIA and Lambda Disclose First Production Co-Packaged Optics Cluster

We covered NVIDIA’s co-packaged optics extensively during the company’s Hot Chips 2025 presentation. NVIDIA’s first version of the packaging is in its Quantum-X switching product, while Spectrum-X will be an updated second-generation part.

NVIDIA Co Packaged Optics With Silicon Photonics At Hot Chips 2025 _Page_13NVIDIA Co Packaged Optics With Silicon Photonics At Hot Chips 2025

NVIDIA is using an array of 18 pluggable light sources on the front of the chassis, as shown in the photos above, to generate light en route to the switch package’s optical modules. Those light sources are pluggable, meaning they can be serviced, which was one of the key inhibitors to CPO in switches when the light source was integrated. Light sources tend to be the parts that fail in pluggable optics, so making them serviceable was an important innovation. Likewise, on the switch package side, making the connections at least somewhat serviceable helps with deployment.

NVIDIA Co Packaged Optics With Silicon Photonics At Hot Chips 2025 _Page_14NVIDIA Co Packaged Optics With Silicon Photonics At Hot Chips 2025

Downstream, NVIDIA’s Quantum-X CPO switch pictured above has 144 MPO/MTP ports, each at 800Gbps. Something interesting is that with only 18 light sources, instead of 144x 800Gbps optical modules (or 72x dual 800G modules), there are actually fewer light sources that can fail than in a traditional switch, and it becomes easier to cool them compared to having a larger array of individual pluggable modules.

One of the biggest benefits of the deployment is that by moving the optics to the switch package rather than the front-panel pluggables, numerous high-speed copper links do not need to drive signaling over relatively long distances from the switch chip to the front cages. As a result, CPO can significantly lower power consumption so that more GPUs can fit in the same footprint. After all, in a power-limited environment, ultimately the more power that can be used for GPUs directly impacts the revenue that can be generated per Watt of power available to the cluster.

Patrick Quebec City Canada November 2026Patrick Quebec City Canada November 2026 in a Yellow STH Tour Filming Shirt

Lambda and NVIDIA may have a short on the cluster up, but let me be clear: it exists and is running today, putting NVIDIA at the forefront of deploying CPO. I think it is actually just being announced this week as part of SC26. As a fun aside, I was at OCP Summit 2026 a few weeks ago when I heard during another vendor’s keynote that CPO was still years away from deployment. It turns out it is already deployed. We no longer have to ask if or when CPO will finally be deployed. The big question now is what the adoption curve will be.

原文为英文,正文由自动抓取并保留原文。

本文由采集器自动同步

原文发布于 ServeTheHome(NVIDIA and Lambda Disclose First Production Co-Packaged Optics Cluster… It is FINALLY real),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗