Nvidia Corp. today disclosed that it has invested $2 billion in Marvell Technology Inc., a publicly traded semiconductor designer.

The cash infusion is part of a new partnership that will span several parts of the chip market.

Santa Clara, California-based Marvell is a major supplier of data center chips. It makes processors optimized to power fiber optic networks and storage hardware. Additionally, Marvell has a business unit that helps other companies develop custom ASICs, or application-specific integrated circuits. That unit is the first focus of its parttnership with Nvidia.

Data center operators link Nvidia graphics processing units to the other chips in their server racks using an interconnect called NVLink. Until recently, the interconnect only worked with the GPU giant’s own chips. Last year, Nvidia introduced a technology called NVLink Fusion that enables custom ASICs such as those made by Marvell to use NVLink.

Marvell will “provide custom XPUs and NVLink Fusion-compatible scale-up networking” to joint customers as part of the partnership. The main component of NVLink Fusion is a chiplet that companies can integrate into their custom processor designs. The chiplet enables the host processor to connect to other chips via NVLink connections.

NVLink is designed to link together chips installed in the same server. Nvidia also offers a second interconnect, NVLink Switch, that can move data across the servers in a rack. NVLink Fusion supports both interconnects.

Nvidia’s partnership with Marvell also extends to other areas. 

The companies will collaborate on Nvidia Aerial, a collection of hardware modules and software tools for powering 5G networks. The product suite lends itself to tasks such as creating a digital twin of a carrier network to find optimization opportunities. In parallel, Nvidia will work with Marvell to “advance world-class networking for AI, including advanced optical interconnect solutions and silicon photonics technology.”

Last year, Marvell acquired an optical interconnect startup called Celestial AI Inc. in a deal worth up to $5.5 billion. The acquisition bought it an optical interposer that can be used to link together a processor’s chiplets. An interposer is a silicon rectangle that functions as a processor’s base layer. It contains tiny wires that transport data between chiplets.

Marvell also competes in other parts of the optical networking market. It sells pluggable coherent transceivers for metro networks, long-distance fiber optic networks that link together remote data centers. Additionally, Marvell makes chips that can be used to move packets between servers installed in the same data center.

“Token generation demand is surging, and the world is racing to build AI factories,” said Nvidia Chief Executive Officer Jensen Huang. “Together with Marvell, we are enabling customers to leverage Nvidia’s AI infrastructure ecosystem and scale to build specialized AI compute.”

The partnership comes a few weeks after Nvidia invested $4 billion in two other optical network equipment suppliers. Lumentum Holdings Inc. and Coherent Corp. make co-packaged optics, or CPO, modules. The technology reduces the number of components necessary to build high-speed fiber optic networks, which lowers the cost of building AI data centers. 

Photo: Marvell