Advertisement


Home Networking NVIDIA and Lambda Disclose First Production Co-Packaged Optics Cluster… It is FINALLY...

NVIDIA and Lambda Disclose First Production Co-Packaged Optics Cluster… It is FINALLY real

0
Lambda NVIDIA Quantum CPO Cover
Lambda NVIDIA Quantum CPO Cover

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_13
NVIDIA 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_14
NVIDIA 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.

Final Words

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 2026
Patrick 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.

LEAVE A REPLY

Please enter your comment!
Please enter your name here

This site uses Akismet to reduce spam. Learn how your comment data is processed.