DriveNets Extends Its Disaggregated AI Fabric With Two Tomahawk 6 Platforms Delivering 102.4 Tb/s for Scale-Up, Scale-Out, and Scale-Across
DriveNets unveiled two Broadcom Tomahawk 6-based AI fabric platforms, the liquid-cooled 2600SL and air-cooled 2601S, each carrying 64 ports of 1.6 Tb/s for 102.4 Tb/s of capacity.
Overview
AI networking vendor DriveNets has expanded its AI fabric portfolio with two high-capacity platforms, the liquid-cooled 2600SL and the air-cooled 2601S, according to SDxCentral. Both systems are built on Broadcom’s Tomahawk 6 chip and feature 64 ports of 1.6 Tb/s to provide a total connectivity capacity of 102.4 Tb/s, and they are set to ship during the current third quarter, SDxCentral reported.
The launch places DriveNets among the first system vendors to build around Broadcom’s flagship switch silicon, which the chipmaker describes as delivering 102.4 terabits per second of switching capacity in a single chip, according to Broadcom.
What We Know
The 2600SL and 2601S are both based on Broadcom’s Tomahawk 6 chip and feature 64 ports of 1.6 Tb/s capacity, adding up to 102.4 Tb/s per system, as reported by SDxCentral. The air-cooled 2601S and liquid-cooled 2600SL are designed to support deployments spanning scale-up inside the rack, scale-out from cabinet to cabinet, and scale-across distributed computing, SDxCentral reported. DriveNets says the 2600SL uses a true 100% liquid-cooled design to maintain thermal efficiency across the whole system and maximize power efficiency, according to SDxCentral.
The platforms extend DriveNets’ distributed disaggregated chassis approach, which spreads a single logical fabric across many standard boxes rather than a single-vendor chassis, as described by SDxCentral. The company has previously provided the core routing system for AT&T among its early wins, SDxCentral noted.
“The most expensive idle asset in the world today is an XPU waiting on the network,” said Ido Susan, DriveNets co-founder and CEO, in remarks reported by SDxCentral. Susan added that the new products “deliver industry-leading performance at massive cluster scale and let our customers maximize infrastructure utilization and power efficiency on any AI accelerator,” and that “these capabilities will only become more critical as the industry moves to heterogeneous AI architectures,” according to SDxCentral.
Charlie Kawwas, president of Broadcom’s Semiconductor Solutions Group, said that “as AI systems scale to hundreds of thousands of XPUs, the network fabric has become a critical factor in AI performance, infrastructure utilization, and overall token economics,” as quoted by SDxCentral.
The Silicon Underneath
Broadcom positions Tomahawk 6 as a chip built for the largest AI clusters. The company says the switch packs 1,024 100G SerDes on a single chip and supports 1.6 Tbps port speeds, according to Broadcom. The chipmaker cites a scale-up cluster size of 512 XPUs and more than 100,000 XPUs in a two-tier scale-out network at 200 Gbps per link, and frames the part as designed for AI clusters with more than one million XPUs, per Broadcom. Tomahawk 6 supports both co-packaged optics and pluggable connectivity, Broadcom said. “Tomahawk 6 is not just an upgrade – it’s a breakthrough,” said Ram Velaga, senior vice president and general manager of Broadcom’s Core Switching Group, according to Broadcom.
DriveNets’ systems arrive amid a broader wave of Tomahawk 6-based hardware; Arista built its recently announced 7060XE7 switch family on the same silicon, as previously reported.
What We Don’t Know
The reporting does not disclose pricing for the 2600SL or 2601S, nor named customers for the new platforms beyond DriveNets’ earlier AT&T routing engagement. Neither source provides independent benchmarks of the platforms’ real-world performance or power consumption, and DriveNets has not detailed availability beyond the third-quarter shipping window.
Analysis
DriveNets’ pitch centers on decoupling AI networking from any single accelerator supplier: by running a distributed fabric across standard Tomahawk 6 boxes, the company argues operators can mix accelerators and still reach cluster-scale performance, a framing reinforced by Susan’s emphasis on running “on any AI accelerator” and on “heterogeneous AI architectures,” according to SDxCentral. The dual air- and liquid-cooled options acknowledge that AI data centers are splitting between dense, liquid-cooled racks and conventional air-cooled halls. With multiple vendors now shipping systems on the same 102.4 Tb/s Broadcom part, competition in AI Ethernet fabrics is increasingly a contest of system design, cooling, and software rather than raw switch bandwidth.