What's Happening?
Broadcom showcased its open Ethernet fabric portfolio at Supercomputing 2026 (SC26) in Chicago, targeting artificial intelligence (AI) and high-performance computing (HPC) environments. The presentation featured the 102.4-terabit Tomahawk 6 switch, Tomahawk Ultra,
Jericho 4, and the 800G Thor Ultra network interface card (NIC) with enhanced RoCE. These technologies offer university research computing centers and national laboratories a vendor-neutral alternative to proprietary interconnect architectures for large-scale GPU and accelerator clusters. The Tomahawk 6 switch, built on a 3-nanometer process, delivers 102.4 Tbps of switching throughput, enabling two-tier, non-blocking leaf-spine fabrics for over 8,192 accelerators at 800Gbps. The Tomahawk Ultra provides sub-microsecond latency for tightly coupled scale-up fabrics, while the Thor Ultra 800G AI Ethernet NIC interfaces with host systems via PCIe Gen6 x16, featuring an enhanced RoCE engine to overcome traditional bottlenecks.
Why It's Important?
This development is crucial for U.S. research and industry, as it addresses the historical reliance on proprietary interconnects like NVIDIA InfiniBand, which created single-vendor dependencies and high costs. Broadcom's open Ethernet solutions aim to democratize access to high-performance networking for AI and HPC, fostering greater competition and innovation. For university clusters and national lab facilities, this means more flexible hardware procurement, allowing them to deploy heterogeneous compute nodes from various vendors across a shared, standards-based Ethernet backbone. The ability to consolidate high-bandwidth workloads onto a unified Ethernet backbone also accelerates research workflows, particularly when transferring large datasets to HPC clusters. Furthermore, the enhanced RoCE engine in Thor Ultra NICs achieves job completion times comparable to InfiniBand quantum fabrics, providing a viable open-standard alternative for demanding AI workloads.
What's Next?
Broadcom is currently sampling its Tomahawk 6 switch silicon and Thor Ultra 800G NICs to primary server OEMs, ODMs, and hyperscale cloud providers. Commercial availability of fixed-configuration leaf-spine switches and liquid-cooled enterprise chassis is anticipated throughout 2027. Concurrently, the Ultra Ethernet Consortium is expected to finalize its 1.0 compliance certification program in early 2027, which will establish verified multi-vendor interoperability for open Ethernet AI infrastructure. This standardization will be critical for widespread adoption and integration into existing and future supercomputing environments. Academic researchers and enterprise cloud engineers will continue to evaluate these open solutions for their potential to simplify data center layouts, reduce transceiver counts, and achieve significant power reductions per gigabit, especially in power-constrained computing centers.
Beyond the Headlines
The shift towards open Ethernet fabrics represents a significant architectural evolution in supercomputing, moving away from specialized, proprietary solutions. This trend aligns with broader community objectives for sustainable HPC and energy proportionality in computing. By enabling standard Linux network stacks, university teams can tune MPI collective routines and AI collective libraries without proprietary lock-in, promoting cross-platform compatibility and reducing operational complexities. The implications extend to supply chain auditing, as standards-compliant Ethernet facilitates firmware verification and provenance, which is increasingly important given federal requirements for export controls and compute governance. This move could also influence the design of future AI silicon, as the emphasis on open standards encourages broader industry participation and innovation in high-performance networking.













