Broadcom has debuted an Ethernet switching chip, the Tomahawk Ultra, designed to enable ultra high-speed networking for AI and HPC workloads using open standards networking technology. The Tomahawk Ultra chip competes with the proprietary NVIDIA NVLink Switch chip, but is capable of connecting four times the processors, according to Broadcom.
As companies race to build high-performing AI infrastructure, there’s enormous focus on next-gen AI chips with high processor speeds, notably NVIDIA’s GPUs. But also crucial to AI performance is the networking gear that connects these chips together into clusters, enabling the computational communication between them that drives the performance of these expensive chips to a far higher level.
This connective technology is known as scale-up computing, and it’s the job the Tomahawk Ultra is designed to do. The Ultra handles AI scale-up clusters up to 256 XPUs, and multi-tier HPC clusters of any size—which is far beyond the requirements for traditional data centers. “With the Tomahawk Ultra, we actually targeted the design for HPC high-performance computing, and that has a different set of requirements than regular sort of hyperscale cloud computing,” said Robin Grindley, principal line manager with Broadcom, in an interview. “If you take an XPU, anything that has Ethernet coming out of it, you can connect it up to Tomahawk Ultra and get the benefit of very low latency and very high packet rate.”
The Ultra, a low latency, 250 nanoseconds Ethernet switch chip, enables the transmission of over 76 billion packets per second (bpps), processing 64B frames at 51.2 terabits per second. It can handle this high level of connectivity, Broadcom claims, without blocking or dropping data transmission. Its Credit-based Flow Control feature boosts cluster performance by enabling exceptionally high traffic loading and connective link utilization. Also improving system performance, its Link Layer Retry feature recovers link errors in hardware data transmission autonomously, which increases uptime.
Broadcom is competing aggressively in the AI infrastructure market. Part of this effort is support for the Ethernet standard, an open standard networking technology that competes with NVIDIA’s proprietary InfiniBand. It has advantages like very low latency and superior performance for tightly coupled HPC and AI workloads. But Ethernet continues to develop, and is favored due to its lower cost and standardization. The advanced tech specs of the Ethernet-based Tomahawk Ultra are clearly an effort by Broadcom to further promote Ethernet in data centers.
“Tomahawk 5 [debuted in 2022] actually started the trend of using Ethernet for that scale-out, connecting these racks together in hundred thousand, million GPU-type clusters,” Grindley said. “And then Tomahawk 6 [debuted in June 2025] accelerates that trend.” As reported by Techstrong.it, Tomahawk 6 was the first switch chip with 102.4 tbps of switching capacity in a single chip, which is double the bandwidth of any Ethernet switch sold today.
Tomahawk Ultra will allow you to use Ethernet for both scale-out and scale-up computing, Grindley said. “You get exactly the same features and performance that you need for scale-up from alternative proprietary solutions.” The advantage of using Ethernet over proprietary solutions, he said, is that “all of the management, all of the system operations side of things can be reused. Any tools that you use for analyzing your network, they work exactly the same, whether it’s in the scale-up domain or the scale-out domain.”
The Tomahawk Ultra, manufactured by Taiwan Semiconductor Manufacturing with its 5 nanometer process, is currently shipping to customers.
