Broadcom has introduced its latest data center networking chip, the Tomahawk Ultra, designed to enhance Ethernet’s performance specifically for high-performance computing (HPC) and artificial intelligence (AI) applications. Unlike traditional Ethernet switches that focus on maximum throughput, the Tomahawk Ultra prioritizes extremely low latency, aiming to provide an edge over Infiniband for scale-up networks.
The Tomahawk Ultra offers a maximum bandwidth of 51.2 Tbps while maintaining a switch latency of just 250 nanoseconds. This is particularly significant for HPC applications, which often require the transfer of small packets of data. The chip is optimized for ultra-low latency communication, allowing it to efficiently handle 64-byte packets—something that previous Ethernet technologies struggled with.
A notable feature of the Tomahawk Ultra is its lossless Ethernet capabilities. It incorporates two advanced features from the Ultra Ethernet Consortium’s 1.0 specification: Link Layer Retry (LLR) and Credit-Based Flow Control (CBFC). LLR enables the chip to automatically detect transmission errors and request retransmissions, while CBFC helps manage buffer overflow situations, thereby reducing packet drops.
Moreover, the chip enhances network efficiency with in-network collectives (INC), which streamline operations involving multiple GPUs. Instead of sending data back and forth for processing, the switch can perform certain computations—like averaging results—within the network infrastructure itself, significantly reducing the amount of data that needs to travel across the network and speeding up overall processing times.
Broadcom has also contributed its Scale-Up Ethernet (SUE) specification to the Open Compute Project, aiming to optimize the communication between various processors—especially GPUs—within a network. This framework supports the use of smaller headers to improve data transfer efficiency, challenging the proprietary technologies like Nvidia’s NVLink and other competing interconnect solutions.
The release of the Tomahawk Ultra marks an essential step in making Ethernet a viable option for intensive networking tasks in HPC and AI environments, allowing for greater flexibility and integration within existing network infrastructures. With these advancements, Broadcom continues to strengthen Ethernet’s position in high-performance networking markets.
