Press release · November 17, 2026 · techblog.comsoc.org
Broadcom showcases open Ethernet fabric for AI and HPC
Summary
Broadcom showed its open Ethernet fabric portfolio for artificial intelligence and high-performance computing at Super Computing 2026, held in Chicago, Illinois, on November 15, 2026. The components were presented as designed to compete with proprietary systems, including Nvidia's InfiniBand and NVLink, in AI and HPC clusters.
Bandwidth now reaches parity across switching ASICs, so competition has shifted toward tail-latency determinism in collective operations. The Ultra Ethernet Consortium identifies manufacturer choice and interoperability as intended benefits, alongside performance, scalability, and total cost of ownership.
The Tomahawk 6 switch provides 102.4 Tbps of switching capacity on a single ASIC, and Broadcom has described architectures using it to connect approximately 128,000 accelerators through two switching tiers. Tomahawk Ultra, rated at 51.2 Tbps, has about 250 ns latency and uses credit-based flow control and link layer retry. Jericho 4 targets inter-pod aggregation and multi-datacenter AI federations.
The Thor Ultra 800G network interface card uses enhanced RDMA over Converged Ethernet, which allows supported RDMA traffic to reach accelerator memory when packets arrive out of order. Packet spraying across multiple paths, selective retransmission, and packet trimming support load distribution and congestion handling, though sends and atomics still require ordering and out-of-order placement is not universal.
Broadcom co-presented technical sessions at the SC26 Workshop on High-Performance Fabrics for AI and HPC, organized with researchers from The Ohio State University. The sessions covered Scale-Up Ethernet Transport under the Open Compute Project and transport standards from the Ultra Ethernet Consortium, which were described as providing deterministic, lossless transmission for collective communication across distributed foundation models.
Switch latency figures are not equivalent to end-to-end network latency or collective completion time, so the approximately 250 ns Tomahawk Ultra figure should not be read as proof of performance parity with proprietary accelerator interconnects. Support for individual UEC features also differs from demonstrated interoperability across a complete multivendor deployment. Workload-performance and energy-efficiency comparisons still require clearly defined test conditions.
Our summary. The original keeps its own byline.