Building for Scale: Expanding the Aria Portfolio with Industry Leading 1.6Tb/s Performance
Over the past year, we’ve written about the generational shift from cost-optimized cloud networking to performance-optimized AI clusters. Since our unveiling at OCP and SC last year, the market response has confirmed what we have been witnessing over the last few years: the decades-old model of networking is colliding with the requirements of AI.
As we’ve stated before, Networks that Think are not enabled by software alone. They require a specialized foundation: optimized hardware, a telemetry-centric software architecture, and domain-native AI models, all working together to meet the extreme requirements of AI clusters.
Today, we are taking a major step forward on the first of those pillars. We’re excited to announce that Aria Networks is expanding our hardware portfolio with 1.6Tb/s systems powered by the Broadcom Tomahawk® 6 (TH6) chipset.
The Power of Interoperable Standards
At Aria Networks, our philosophy is rooted in the power of Standard Ethernet. We believe that the future of AI cannot be built on proprietary islands. By doubling down on interoperable protocols, we ensure that the Aria portfolio works seamlessly across every layer of the data center—from scale-out and scale-up fabrics to scale-across and front-end networks.
Crucially, our architecture is accelerator-agnostic. Whether you are deploying NVIDIA GPUs, Google TPUs, AMD accelerators, or specialized silicon like Cerebras, Aria provides a consistent, high-performance fabric. This flexibility ensures that your networking investment remains a long-term asset, regardless of how the AI chip landscape evolves.
A Modular Approach to Massive Scale
To realize the potential of 100,000+ GPU clusters, we’ve adopted a modular architectural approach. Rather than building overly complex, monolithic systems, we scale the network by utilizing hundreds—if not thousands—of the same high-performance components.
This modularity is the key to reliability and performance at scale. The true differentiation isn’t just in the hardware modules themselves; it’s in how they are combined into one cohesive system. We manage this massive, distributed fleet through our unified software layer that optimizes performance and guarantees reliability. This is where the benefits of being “telemetry-first” come to life - by ingesting fine-grained data from every module in the fabric, the NICs, and the hosts, our specialized AI layer can optimize the networking domain as one system.
Leading with Liquid Cooling (200G SerDes)
As power density continues to climb, air cooling is increasingly a bottleneck. For many next-generation AI deployments, liquid cooling is the only viable path to density and efficiency.
At Aria, we are leading in this transition. We will begin shipping our 2U, 64-port OSFP 1.6Tb/s liquid-cooled system to initial customers in February. While much of the industry is still in the planning phase, our partners will be the first to validate 200G liquid-cooled deployments in production-relevant environments.
We want to be clear about the reality of the market: high-end networking silicon remains in extremely high demand. We are currently making allocations for our initial 200G Tomahawk 6 run. If your roadmap includes this transition, early engagement is required to ensure your project stays on schedule.
Optimizing for the Real World: 128 x 800G Without Gearboxes
While 200G SerDes represents the leading edge, the reality on the ground is that the majority of deployed AI NICs still operate at 100G SerDes. As the industry transitions to 200GbE, many architectures rely on “gearboxes” to bridge that gap—adding cost, latency, and power consumption.
We recognize this transitional state. We expect to be the first vendor to deliver a 128 x 800GbE Tomahawk 6 system based on a native 1024 x 100G configuration. This approach enables:
-
Gearbox-Free Efficiency: Native 100G-to-100G connectivity that removes unnecessary latency.
-
Massive Scale: Double the radix of Tomahawk 5, enabling clusters of up to 32,000 accelerators in a simple, high-performance two-tier topology.
Aria Networks Portfolio Overview
| Feature | Aria TH5 | Aria Tomahawk 6 (High Radix) | Aria Tomahawk 6 (Liquid) | Aria Tomahawk 6 (Air) |
|---|---|---|---|---|
| Throughput | 51.2 Tbps | 102.4 Tbps | 102.4 Tbps | 102.4 Tbps |
| Max Port Speed | 800 GbE | 800 GbE | 1.6 TbE | 1.6 TbE |
| Ports | 64x 800G | 128x 800G | 64x 1.6T | 64x 1.6T |
| SerDes | 512 x 100G | 1,024 x 100G | 512 x 200G | 512 x 200G |
| Cooling | Air (2U) | Air (4U) | Liquid (2U) | Air (4U) |
| Strategic Focus | Proven AI Scale-out | 32k GPU Radix / No Gearbox | Max Density & Efficiency | High-speed Air-cooled |
White Glove Alignment
As with all our offerings, these systems come with the Aria “white glove” treatment. Our Forward Deployed Engineers (FDEs) act as an extension of your networking team, managing the complexity of these massive, modular fabrics from A to Z.
We are building the networking foundation required for the next decade of AI scale. If you are designing clusters where performance, interoperability, and scale truly matter, reach out to sales@arianetworks.com to discuss your roadmap and secure your hardware allocations.
The journey to Networks that Think is accelerating. We’re just getting started.
Supporting Quotes
Hasan Siraj, Vice President of Product Marketing, Core Switching Group, Broadcom
“Proprietary fabrics are a thing of the past. With its 1.6 Tbps launch, combined with a telemetry-centric software architecture, Aria Networks is proving that the highest-performance AI networks on the planet are being built on a foundation of open, scalable Ethernet such as Broadcom’s Tomahawk 6 switch series.”
Sameh Boujelbene, Vice President at Dell’Oro Group
“The AI era is driving a fundamental shift in data center networking, with 1.6T Ethernet emerging as the new industry standard. As AI back-end networks demand unprecedented scale, performance, and real-time visibility, open and interoperable architectures are no longer optional. Aria Networks’ portfolio directly addresses this industry inflection by pairing 102.4 Tbps standard Ethernet with advanced telemetry to deliver the scale, performance, and openness required for the next generation of AI infrastructure.”
Dylan Patel, Founder of SemiAnalysis
“With the launch of its 102.4Tbps portfolio, Aria is solving the AI connectivity bottleneck. By pairing the scale of standard Ethernet with the raw performance of Broadcom’s Tomahawk 6 switch and the power of its telemetry-centric software architecture, Aria is providing the definitive blueprint for the next generation of AI clusters.”
