On November 15, 2026, Broadcom showcased its open Ethernet fabric portfolio for artificial intelligence and high-performance computing at SC26 in Chicago. Featuring the 102.4-terabit Tomahawk 6 switch, Tomahawk Ultra, Jericho 4, and the 800G Thor Ultra network interface card with enhanced RoCE, the demonstration provides university research computing centers and national laboratories an open, vendor-neutral alternative to proprietary interconnect architectures for large-scale GPU and accelerator clusters.
Open Ethernet Silicon Showcased at SC26
At the SC26 supercomputing conference held at McCormick Place in Chicago (November 15–20, 2026), Broadcom Inc. demonstrated a unified, open Ethernet networking portfolio targeting hyperscale artificial intelligence (AI) and high-performance computing (HPC) environments. The showcase centered on the newly unveiled StrataXGS Tomahawk 6 (BCM78910 series) switch silicon, the Tomahawk Ultra low-latency scale-up switch, and the Thor Ultra 800-gigabit Ethernet (800GbE) network interface card (NIC) featuring enhanced RDMA over Converged Ethernet (eRoCE). Paired with the deep-buffer Jericho 4 fabric router, the silicon suite demonstrates how open Ethernet now addresses both intra-node scale-up accelerator communication and multi-thousand-node scale-out fabrics without requiring proprietary interconnect fabrics.
Broadcom also co-presented technical sessions at the SC26 Workshop on High-Performance Fabrics for AI and HPC, organized in collaboration with researchers from The Ohio State University. The technical presentations highlighted how emerging specifications—including Scale-Up Ethernet Transport (SUE-T) under the Open Compute Project and transport standards from the Ultra Ethernet Consortium (UEC)—provide deterministic, lossless transmission for collective communication routines across distributed foundation models.
Technical Specifications: Tomahawk 6, Tomahawk Ultra, and Thor Ultra
The silicon demonstrated in Chicago addresses critical bandwidth and latency constraints in large-scale accelerator deployments. Chief among the announcements is Tomahawk 6, fabricated on an advanced 3-nanometer process to deliver 102.4 terabits per second (Tbps) of switching throughput on a single monolithic ASIC. The device integrates 512 physical SerDes channels running at 212.5 Gbps PAM4, enabling radix options such as 64 ports of 1.6-terabit Ethernet (1.6TbE), 128 ports of 800GbE, or 512 ports of 200GbE. By doubling the switching capacity of prior 51.2T devices, Tomahawk 6 enables two-tier, non-blocking leaf-spine fabrics supporting more than 8,192 accelerators at 800Gbps without intermediate aggregation stages.
For tightly coupled scale-up fabrics within accelerator domains, Broadcom demonstrated Tomahawk Ultra, a 51.2 Tbps switch ASIC optimized for sub-microsecond latency. Tomahawk Ultra achieves approximately 250 nanoseconds of port-to-port cut-through latency utilizing hardware Credit-Based Flow Control (CBFC) and Link Layer Retry (LLR), matching proprietary interconnect latencies while retaining standard IEEE 802.3 frames.
At the server endpoint, the Thor Ultra 800G AI Ethernet NIC interfaces with host systems via PCIe Gen6 x16. Thor Ultra incorporates an eRoCE engine designed to resolve traditional RoCEv2 bottlenecks. It features hardware out-of-order packet reassembly directly into GPU memory via kernel dma-buf paths, packet spraying, hardware selective acknowledgments (SACK), and programmable congestion signaling (CSIG) with packet trimming, maintaining high fabric utilization under heavy traffic loads.
| Component | Silicon Class | Throughput | Key Architectural Features | Target Topology Role |
|---|---|---|---|---|
| Tomahawk 6 (BCM78910) | Ethernet Switch ASIC | 102.4 Tbps | 3nm process, 512 × 200G SerDes, dynamic load balancing, UEC telemetry | Scale-out spine/leaf fabrics for massive GPU clusters |
| Tomahawk Ultra | Low-Latency Switch ASIC | 51.2 Tbps | ~250ns latency, Credit-Based Flow Control (CBFC), Link Layer Retry (LLR) | Intra-rack scale-up accelerator memory fabrics |
| Thor Ultra | AI Network Interface Card | 800 Gbps (2 × 400G / 1 × 800G) | PCIe Gen6 x16, eRoCE, direct GPU memory placement, packet trimming | Host accelerator endpoints (GPUs, TPUs, XPUs) |
| Jericho 4 | Fabric / Core Router ASIC | Multi-Tbps Modular | Deep HBM packet buffers, virtual output queuing (VOQ) | Inter-pod aggregation and multi-datacenter AI federations |
Architectural Evolution: Moving Beyond Proprietary Interconnects
Historically, high-performance computing centers and distributed AI clusters relied on proprietary fabrics—primarily NVIDIA InfiniBand and proprietary NVLink networks—to achieve low latency and lossless communication. While effective for traditional MPI applications, proprietary fabrics created single-vendor supply dependencies and steep cost premiums that challenged institutional IT budgets. Early attempts to deploy commodity RoCEv2 struggled under AI workloads because Priority-based Flow Control (PFC) was prone to head-of-line blocking and slow Go-Back-N packet recovery.
Broadcom’s showcase indicates that open Ethernet has resolved these issues. Technical analysis by ServeTheHome notes that Tomahawk 6 switches combined with Thor Ultra eRoCE NICs achieve job completion times matching 800G InfiniBand quantum fabrics. Because Thor Ultra places packets directly into GPU memory regardless of arrival order, switches can spray packets across all available links. When congestion occurs, switches trim packet payloads while forwarding headers, triggering instantaneous retransmissions. This architectural shift aligns with standards established by the Ultra Ethernet Consortium to support clusters scaling beyond 100,000 accelerators.
Industry and Academic Reaction
The demonstrations at SC26 generated significant engagement among academic supercomputing center directors and enterprise cloud engineers. Cloud infrastructure architects noted that 102.4T single-chip capacity substantially simplifies data center physical layouts by reducing transceiver counts and cabling bulk. Consolidating switching capacity onto 102.4T silicon yields estimated per-gigabit power reductions of up to 40 percent compared to dual-chassis 51.2T configurations, an essential efficiency gain for power-constrained university computing centers.
Academic researchers participating in the High-Performance Fabrics workshop emphasized the strategic advantage of open networking ecosystems. Standardizing on open Ethernet interfaces enables university teams to tune MPI collective routines and AI collective libraries (such as NCCL and RCCL) using standard Linux network stacks, avoiding proprietary lock-in while maintaining cross-platform compatibility.
Implications for University Clusters and National Lab Facilities
The availability of 102.4T switching and 800G open Ethernet silicon introduces several strategic benefits for university research computing facilities and sponsored research offices:
- Vendor-Neutral Hardware Procurement: Open Ethernet decouples accelerator procurement from network fabrics. Campus research computing teams can deploy heterogeneous compute nodes from multiple hardware vendors across a shared, standards-based Ethernet backbone.
- Converged Data and Storage Infrastructure: Rather than maintaining separate fabrics for compute interconnects and storage networks, institutions can consolidate high-bandwidth workloads onto a unified 100T/800G Ethernet backbone. This setup accelerates workflows when transferring large research datasets to HPC clusters.
- Predictable Multi-Tenant Slurm Scheduling: University clusters shared by diverse research groups benefit from hardware-level dynamic load balancing and packet trimming. Investigators requesting GPU resources in Slurm batch jobs experience consistent runtimes and reduced communication variance during concurrent job runs.
- Alignment with Federal Cyberinfrastructure: National resources allocated through NSF ACCESS compute allocations and major leadership-class systems like the El Capitan supercomputer prioritize open standards and energy efficiency. Broadcom’s support for Co-Packaged Optics (CPO) aligns with federal initiatives targeting sustainable high-performance computing.
- Transparent Supply Chain Auditing: As institutions manage federal requirements governing export controls and compute governance thresholds, standards-compliant Ethernet facilitates firmware verification and supply chain provenance.
Roadmap and Deployment Timelines
Broadcom confirmed that Tomahawk 6 switch silicon and Thor Ultra 800G NICs are currently sampling to primary server OEMs, ODMs, and hyperscale cloud providers. Commercial fixed-configuration leaf-spine switches and liquid-cooled enterprise chassis are expected to reach general availability across 2027. Concurrently, the Ultra Ethernet Consortium is scheduled to finalize its 1.0 compliance certification program in early 2027, establishing verified multi-vendor interoperability across open Ethernet AI infrastructure.
Sources
- Broadcom Inc., ‘Broadcom Showcases Open Ethernet Infrastructure for AI and HPC at SC26’, Broadcom Product Announcements (2026-11-15).
- ServeTheHome, ‘Broadcom Tomahawk 6 102.4T and Thor Ultra 800G eRoCE Ethernet Silicon Analysis’, ServeTheHome Enterprise Networking (2026-11-15).
- SC26 Supercomputing Conference, ‘SC26 Technical Program and Exhibition Guide: High-Performance Networking in Chicago’, ACM/IEEE-CS (2026-11-15).
- Ultra Ethernet Consortium, ‘Ultra Ethernet Consortium Architecture and Scale-Up Specifications Overview’, Linux Foundation / UEC (2026-11-15).
- The Ohio State University Network-Based Computing Laboratory, ‘Workshop on High-Performance Fabrics for AI and HPC at SC26’, MVAPICH Project Proceedings (2026-11-15).







