Broadcom has announced the launch of its Tomahawk Ultra chip, a high-bandwidth Ethernet switch ASIC based on the company’s proprietary Scale-Up Ethernet (SUE) architecture. The launch comes hot on the heels of its Tomahawk 6 announcement in June.

SUE vs NVLink:

SUE is designed for hyperscale data centers developing rackscale AI training clusters using Ethernet-based AI fabrics. Unlike NVIDIA’s NVLink, SUE does not directly connect GPUs to GPUs within a node or server and does not have memory coherency. It is therefore not a direct alternative to compute fabrics such as NVLink, CXL, PCI Gen5/Gen6, or the upcoming UALink standard, which are all point-to-point direct interconnects designed specifically for GPU-to-GPU communications. Instead, SUE is an Ethernet-based technology which operates at server level via a Network Interface Card (NIC).

AI Training:

According to Broadcom, SUE is designed for distributed AI training for applications with moderately high interconnect bandwidth needs. However, even with its various enhancements, it still has higher latency and lower bandwidth than NVIDIA’s NVLink. As a result, it cannot handle ultra-tight, low-latency synchronization across thousands of GPUs. For example, it is suitable for training 100B+ parameter, transformer-based models with dense model parallelism but cannot handle 100B-1T+ parameter models such as GPT-4, and Gemini.

Ultra Ethernet vs SUE:

Unlike the Ultra Ethernet Consortium (UEC) specification – a fully open Ethernet standard – Broadcom’s SUE is vendor-specific with customized features that are optimized for performance. However, it comes at a cost of limiting customers to using Broadcom silicon and software stacks. The Tomahawk Ultra ASIC is sampling now, and the first commercial products are expected in H1 2026.

Analyst Take:

Broadcom’s SUE is not a direct replacement for NVLink, and whether it is a threat or not really depends on the use case. It could be a threat for AI training where ultra-tight GPU coupling is less important and where customers prefer open standards. However, for tightly coupled AI workloads deeply integrated with NVIDIA’s CUDA, NVIDIA will reign supreme.
Broadcom’s SUE could also be a challenge to NVIDIA’s Infiniband and Spectrum-X scale out technologies, particularly for non-NVIDIA accelerators. Unlike NVIDIA, Broadcom’s SUE combines scale-up and scale-out in a single networking technology. NVIDIA will likely dominate in ultra-low latency, multi-mode training in CUDA-heavy clusters.

Author