Verity 6.6: Built for the East-West Era
When we started scoping Verity 6.6, the customer conversations kept landing in the same place: the network is no longer dominated by north-south flows. AI training clusters, distributed inference, and high-density GPU pods have pulled the center of gravity east-west — and the tools we use to design, deploy, and operate the fabric have to follow.
Verity 6.6 is that follow-through. It’s available today.
Here’s what shipped, and why each piece is in the release.
Designing for AI fabrics
Spectrum-X Reference Architecture
We’re shipping a validated reference design aligned to NVIDIA’s Spectrum-X blueprint, including adaptive routing, congestion control, and the topology shape needed to keep GPU-to-GPU collectives off the slow path. If you’re standing up a Spectrum-X pod, Verity will instantiate the design, not just describe it: switch roles, link maps, fabric-wide policy, and telemetry hooks come out of the box.
Cumulus NOS Support
NVIDIA Cumulus Linux is now a first-class NOS in Verity, alongside the platforms we already manage. Same intent model, same workflows, same validation, point Verity at your Cumulus switches and they participate in the fabric like any other node. This was the most-requested NOS in our last customer survey, and it unlocks much of the Spectrum-X work above.
Spine Planes
Modern AI fabrics aren’t single-spine; they’re plane-based. Each leaf connects to N parallel spines, and traffic is hashed or steered across planes for predictable bisection bandwidth. Verity 6.6 models spine planes natively, so you can express rail-optimized or fully-rail-connected topologies without hand-rolling the wiring map.
East-West Topology Template
A new topology template purpose-built for east-west designs. Leaf/spine ratios, oversubscription targets, plane counts, and rail assignment are all parameterized.
GPU Cabling Topology Validation
The boring one that quietly saves projects. Miscabled GPUs are the silent killer of cluster performance, everything looks healthy until a collective is unexpectedly bottlenecked on one rail. Verity now validates GPU-to-switch cabling against the intended topology and flags misconnects before they go into production.
East-West NIC Management
The fabric doesn’t stop at the ToR. 6.6 brings NICs on the GPU hosts into Verity’s scope: NIC roles, port assignments, and QoS configuration are managed as part of the fabric intent rather than as an afterthought on the server team’s runbook.
Scale
Multifabric Support
A single Verity instance can now manage multiple fabrics. Practically, that means training, inference, and storage fabrics – or per-site fabrics – running under one control plane, with isolation where you need it and shared policy where it makes sense. For customers operating several AI pods, this collapses what used to be N Verity deployments into one.
Day-2 operations
The deployment is the easy part. Most of the lifecycle cost of a fabric is what happens after it’s running, so 6.6 invests heavily in operational visibility and control.
DHCP with MAC Binding
DHCP leases can now be bound to specific MAC addresses inside Verity-managed scopes. Useful for keeping management addresses stable, useful for compliance environments where ad-hoc IP assignment isn’t acceptable, and useful for the engineer trying to track down which NIC has which lease at 2 a.m.
Dell OME Dashboards
For customers standardized on Dell, we now integrate with OpenManage Enterprise. Hardware health, firmware state, and inventory from OME surface alongside Verity’s fabric telemetry, one place to look when something’s wrong.
ARP/ND Cache Workbench
This one’s overdue. Diagnosing ARP and IPv6 ND issues across a large fabric historically meant SSH’ing into N switches and diffing tables by hand. The new workbench pulls cache state from every node, lets you query and compare across them, and highlights inconsistencies. It’s the tool we wish we’d had on every escalation for the last two years.
Upgrading
6.6 is a non-disruptive upgrade from 6.5.x. Release notes and the upgrade guide are on the customer portal. If you’re running a multi-site deployment and want help sequencing the rollout, your account team has the playbook.
A short note on what comes next: 6.6 closes out the AI-fabric design and visibility story we started in 6.4. The next release shifts toward closed-loop performance, telemetry-driven rebalancing, automated PFC tuning, and tighter integration with workload schedulers. More on that soon.
If you’d like a walkthrough of any of the features above, we’re running deeper sessions over the next few weeks. Reach out to your account team or drop a note via the contact form.
جوش سول
Senior Vice President – Product
Josh Saul has pioneered open source network solutions for more than 25 years. As an architect, he built core networks for GE, Pfizer and NBC Universal. As an engineer at Cisco, Josh advised customers in the Fortune 100 financial sector and evangelized new technologies to customers. More recently, Josh led marketing and product teams at VMware (acquired by Broadcom), Cumulus Networks (acquired by NVIDIA), and Apstra (acquired by Juniper).
