Nvidia is betting on heterogeneity as the next phase of enterprise AI.
To provide customers more choice and flexibility, the chip giant has announced a new partnership with Marvell Technology that will connect the semiconductor company to the Nvidia AI factory and AI-RAN ecosystems.
Nvidia is also investing $2 billion in Marvell, and the two companies will collaborate on next-gen 5G/6G networks that support AI workloads.
Nvidia seems to be inking new deals daily, and partnering with companies serving every level of the stack. But analysts view the Marvell partnership as a strategic move based on enterprise demand.
“What we see from Nvidia signals something necessary in the market: A universal control layer that connects and manages the heterogeneous enterprise AI environment,” said Matt Kimball, VP and principal analyst at Moor Insights & Strategy. “Heterogeneity is a destination for virtually every enterprise.”
Key to winning the AI race
Through the new partnership, Marvell will provide custom XPUs (specialized processing units) and scale-up networking that is compatible with Nvidia’s NVLink Fusion rack-scale platform. Customers can now develop “semi-custom” AI infrastructure using Nvidia NVLink, and integrate with Nvidia’s GPUs, large processing units (LPUs), and networking and storage platforms. These include the Vera CPU, ConnectX NICs, Bluefield data processing units (DPUs), NVLink interconnect and Spectrum-X switches, and rack-scale AI compute.
Enterprises building out a Nvidia-connected AI environment will now be able to deploy non-Nvidia accelerators into that environment, Kimball explained. Thus, semi-custom silicon can integrate “much more directly” into Nvidia-based AI systems.
With this move, Nvidia is “further acknowledging the heterogeneity that will be the AI inference environment,” noted Kimball.
Yaz Palanichamy, senior advisory analyst at Info-Tech Research Group, agreed that welcoming Marvell into its NVLink ecosystem increases Nvidia’s support of semi-custom and heterogeneous architectures while allowing customers to continue using its platform. “Enterprise customers [will have] more flexibility when creating their AI systems, but will still create a larger presence in the greater AI ecosystem for Nvidia,” he said.
As Kimball further pointed out, even if Nvidia is the dominant chip in an enterprise’s infrastructure, there will be use cases and deployment scenarios in which third-party chips are required. So the key is to control the fabric and software that ties this heterogenous environment together, which is what Nvidia is aiming for.
There is a “battle of sorts” going on, he noted. While NVLink delivers a high-performance interconnect for Nvidia environments, the competing Ultra Accelerator Link (UALink) is a consortium-based spec that delivers the same capability and is backed by the likes of Astera Labs, AMD, Intel, Meta, Broadcom, and Marvell itself.
“Openness-ubiquitousness is the real key to winning,” said Kimball. “Nvidia is working to shift from a proprietary to a ubiquitous model.”
What this means for enterprises
The partnership will likely only benefit most enterprises indirectly, analysts note.
The vast majority of enterprise IT orgs aren’t going to be building NVLink- or UALink-based systems, Kimball emphasized. Many are already being enabled by cloud providers or original equipment manufacturers (OEMs) in this area.
However, he pointed out, Nvidia is the market share leader, and NVLink being enabled by more and more chip companies such as Intel and Marvell makes it easier to deploy and support in the enterprise. UALink, however, does not support Nvidia chips, ultimately limiting its effectiveness.
Nvidia is continuing to invest billions up and down the stack, and is inking partnerships with the biggest players in the business, including $2 billion for Marvell, a combined $4 billion for photonic companies Coherent and Lumentum to accelerate AI optics and networking technology, and the purchase of $5 billion in Intel stock to connect Nvidia and Intel Xeon architectures using NVLink.
The company is clearly investing to build the ecosystem around its GPUs, LPUs and silicon portfolio, Kimball noted. It’s a smart move, he added. It expands “ecosystem alignment” while improving overall performance in the Nvidia-powered enterprise.
“If successful, Nvidia will increase the diversity of silicon as it consolidates control around the fabric and software that ties it all together,” Kimball noted.
Building out telecom networks to support AI
Beyond the AI stack itself, Nvidia and Marvell will work together to adapt existing telecom infrastructure AI using Nvidia’s Aerial AI-RAN for 5G and 6G networking. They will also work to enhance advanced optical interconnect solutions (to transmit signals across long distances) and silicon photonics technology (using light as opposed to electrical signals to transfer data between computer chips).
“As AI becomes more integrated into the telecom network infrastructure, it will help speed up the rollout of AI-RAN and enable better edge inference,” noted Info-Tech’s Palanichamy. This is critical for the new real-time applications being developed, he added, and as inference becomes more and more distributed.
Ultimately, the move indicates that Nvidia isn’t just thinking about the data center in the traditional sense, Kimball noted. “It’s thinking about a more distributed AI fabric that extends into carrier networks.”
To date, Nvidia has been very datacenter-centric, but it is moving more to the edge, he noted. As this happens, the network is not just a transport, but a core part of the compute architecture.
“If a telecom network evolves into AI-powered infrastructure, it can have real implications for latency-sensitive and operational AI use cases,” said Kimball.