Cisco launches Nexus HyperFabric AI clusters, a GenAI stack for on-prem

Cisco launches Nexus HyperFabric AI clusters, a GenAI stack for on-prem

Cisco is turning previously announced collaboration with Nvidia into a full stack that organizations can use to run GenAI workloads. The new stack goes by the name Cisco Nexus HyperFabric AI clusters.

Earlier this year, we wrote about a collaboration between Cisco and Nvidia. In itself, that was not so special, since everyone is working with Nvidia these days. One of the parts of the collaboration back then was about so-called validated designs that the two vendors set up. These make it possible to run AI workloads optimally, is the idea. In addition, Cisco announced that it was going to offer Nvidia GPUs in UCS rack and blade servers and was going to put Nvidia’s AI software on its own price list.

Cisco Nexus HyperFabric AI clusters

With Cisco Nexus HyperFabric AI clusters, however, the company goes a lot further in its partnership with Nvidia. This is a complete turn-key AI stack. It is a combination of Cisco’s network (i.e., Ethernet-based) portfolio and Nvidia’s accelerated compute components and software. The data portion of Cisco Nexus AI clusters is provided by VAST data. That vendor provides the unified data platform.

Cisco Nexus HyperFabric AI clusters should allow organizations to set up a full AI stack in the private cloud, or on-prem environment. The idea is that from here, organizations can not only design and deploy their AI stack, but also monitor and manage it. The outcome should be an AI infrastructure guaranteed to work well in enterprise environments. In addition to performance, simplicity also plays an important role. Organizations can roll out AI infrastructure across the entire distributed IT environment thanks to central management capabilities.

The components of Cisco Nexus HyperFabric AI clusters

A full AI stack is a rather complex collection of components. Cisco Nexus HyperFabric AI clusters therefore consist of quite a few different components to bring compute, network and storage together. Below we list these.

  • Cisco Nexus 6000 series spine/leaf switches for the Ethernet backbone. These can deliver 400G and 800G bandwidths. Cisco’s ASICS division plays an important role in this, developing the Silicon One offering.
  • To achieve the above bandwidths, the new offering uses Cisco’s own QSFP-DD optical transceivers. These are available in 400G and 800G variants.
  • Manage from the cloud via the Nexus dashboard and Cisco Intersight.
  • Nvidia provides the AI Enterprise software. Part of this is the NIM Microservices that the company recently announced. These are preconfigured and populated containers of inferencing microservices. These should speed up the rollout of AI in organizations.
  • Nvidia Tensor Core GPUs, starting with the H200 NVL.
  • AI is not complete without talking about DPUs these days. For Cisco Nexus HyperFabric AI clusters, Nvidia’s BlueField-3 DPU was chosen. There is also a BlueField-3 SuperNIC to further accelerate the network.
  • All this is built around the Nvidia MGX reference design for servers.
  • Finally, as already indicated, VAST data’s unified data platform has been used.

Cisco Nexus HyperFabric AI clusters will be available to selected Cisco customers beginning in Q4 of Cisco’s fiscal year, with general availability to follow soon after. Jonathan Davidson, EVP and GM Cisco Networking also indicates that the plan is to get storage companies like Pure Storage and NetApp into the HyperFabric ecosystem too. So there’s a lot more to come.