VIAVI Solutions introduced what it describes as the industry’s first Ultra Ethernet Transport (UET) validation solution for AI fabrics, extending its VIAVI TestCenter platform with traffic generation and analysis capabilities designed for next-generation AI data center networks. The new offering targets hyperscalers, cloud providers, neocloud operators, and network equipment vendors developing Ultra Ethernet-based AI infrastructure.
The platform enables validation of both scale-up and scale-out AI networks without requiring physical GPU clusters. Instead, it emulates realistic AI traffic patterns at the transport layer, allowing operators to test network performance, resiliency, congestion management, and load balancing before deploying large-scale AI environments. The announcement comes as the Ultra Ethernet Consortium’s UEC 1.0 specification gains traction as an Ethernet-based alternative for connecting massive AI clusters.
VIAVI said its solution supports full-fidelity emulation of AI workloads, including collective communications (CCL) and large language model (LLM) traffic flows. The platform can generate realistic UET traffic patterns, including Reliable Ordered Delivery (ROD), Reliable Unordered Delivery (RUD), packet trimming, congestion control, and dynamic multipathing. It also validates AI fabric load-balancing mechanisms such as ECMP, packet spraying, and flowlet switching. VIAVI highlighted collaboration with HPE and Juniper Networks, including validation of UET transport on the Juniper QFX5240 platform running Junos Evolved software.
• GPU-free validation platform for Ultra Ethernet-based AI fabrics
• Supports emulation of realistic AI workload traffic, including CCL and LLM flows
• Validates Ultra Ethernet Transport (UET) features such as ROD, RUD, packet trimming, congestion control, and dynamic multipathing
• Enables testing of AI fabric load balancing including ECMP, packet spraying, and flowlet switching
• Designed for hyperscalers, cloud providers, neocloud operators, and network equipment vendors
• Supports validation of both scale-up and scale-out AI network architectures
• Collaboration includes testing on the Juniper QFX5240 running Junos Evolved
“AI clusters will soon scale to millions of endpoints, which means relying on physical GPUs alone to validate network behavior is no longer practical,” said Aniket Khosla, Vice President of Product Management, Optical Transport and High-Speed Ethernet at VIAVI Solutions. “The launch of this GPU-free, full-fidelity UET validation solution gives customers the confidence to deploy scalable, high-performance AI fabrics faster and more cost-effectively.”
🌐 Analysis
The significance of this announcement extends beyond test equipment. As AI clusters scale toward hundreds of thousands—and eventually millions—of endpoints, network architects increasingly need methods to validate transport behavior without assembling expensive GPU infrastructure. This creates a growing market for high-fidelity emulation platforms capable of reproducing AI traffic patterns while exposing congestion, latency, and load-balancing issues before production deployment.
The launch also highlights growing momentum behind the Ultra Ethernet Consortium and its effort to establish Ethernet as a primary interconnect technology for large-scale AI systems. While proprietary approaches such as NVIDIA InfiniBand remain dominant in many AI deployments, the broader industry—including switch vendors, hyperscalers, and networking suppliers—continues to invest in open Ethernet-based AI fabrics. Validation tools capable of emulating UET traffic will likely become increasingly important as vendors move from standards development to large-scale production deployments.
| Ultra Ethernet (UEC) Profile | |
|---|---|
| Updated: June 2026 | |
| Organization | Ultra Ethernet Consortium (UEC) |
| Announced | July 2023 |
| Specification Status | Ultra Ethernet Specification v1.0 released in June 2025 |
| Mission | Create an open Ethernet-based communications stack optimized for AI and High Performance Computing (HPC) clusters. |
| Primary Goal | Modernize Ethernet for AI-scale fabrics while preserving Ethernet/IP ecosystem economics, interoperability, and supply-chain breadth. |
| Transport Layer | Ultra Ethernet Transport (UET), a transport protocol designed specifically for AI and HPC workloads. |
| Key Innovation | Purpose-built transport architecture that optimizes collective communications, congestion handling, multipathing, and network utilization for distributed AI training. |
| Delivery Modes | ROD (Reliable Ordered Delivery), RUD (Reliable Unordered Delivery), RUDI (Reliable Unordered Delivery for Idempotent Operations), and UUD (Unreliable Unordered Delivery). |
| Why Delivery Modes Matter | By relaxing strict ordering requirements where applications permit, UET reduces buffering requirements, minimizes endpoint state, enables packet spraying, and improves utilization across multiple network paths. |
| In-Network Compute (INC) | Supports switch-based collective operation acceleration, including AllReduce offload and other in-network compute capabilities that can reduce training communication overhead. |
| Congestion Management | Advanced congestion control, packet trimming, path-aware forwarding, telemetry, dynamic multipathing, and adaptive traffic steering. |
| Load Balancing | ECMP, packet spraying, adaptive routing, flowlet switching, and dynamic path optimization. |
| AI Workloads | Large Language Models (LLMs), collective communications libraries, distributed training, inference clusters, HPC simulations, MPI workloads, and RDMA-style communications. |
| Software Integration | Designed to support existing AI and HPC software frameworks and communication libraries, including libfabric-based environments. |
| Scale Target | Designed for clusters ranging from thousands to more than one million endpoints. |
| Ethernet Speed Roadmap | 400GbE, 800GbE, and future 1.6TbE deployments aligned with the broader Ethernet ecosystem roadmap. |
| Physical Infrastructure | Leverages existing Ethernet switches, NICs, optics, cabling, routing infrastructure, and management frameworks. |
| Security | Includes scalable security concepts intended for very large AI clusters and multi-tenant environments. |
| Founding Members | AMD, Arista Networks, Broadcom, Cisco, Eviden, HPE, Intel, Juniper Networks, Meta, and Microsoft. |
| Ecosystem Participants | Hyperscalers, cloud providers, neocloud operators, switch vendors, NIC suppliers, optical vendors, server manufacturers, software vendors, and test equipment providers. |
| Competitive Context | The industry’s primary open alternative to NVIDIA InfiniBand for large-scale AI fabrics. |
| Advantages | Open ecosystem, multi-vendor interoperability, Ethernet economics, broad supply chain support, and alignment with cloud networking architectures. |
| 2026 Milestones | Expansion of UET-enabled hardware, software stacks, interoperability testing, switch implementations, NIC support, and validation platforms such as VIAVI TestCenter. |
| Long-Term Vision | Establish Ethernet as the dominant open networking fabric for AI training, inference, HPC, and exascale computing infrastructure. |
