Broadcom delivers Ethernet switch for HPC and AI

Broadcom Inc. has launched the Tomahawk Ultra Ethernet switch for high-performance computing (HPC) and AI workloads. Designed for HPC environments and AI clusters, the Tomahawk Ultra delivers ultra-low latency, high throughput, and lossless networking.

The Tomahawk Ultra delivers ultra-low latency of 250 ns at a full 51.2 Tbits/s throughput. It offers line-rate switching performance even at minimum packet sizes of 64 bytes, supporting up to 77 billion packets per second.

Broadcom's Tomahawk Ultra Ethernet switch.

(Source: Broadcom Inc.)

The Ethernet switch uses lossless fabric technology—Link Layer Retry (LLR) and Credit-Based Flow Control (CBFC)—to eliminate packet loss. LLR enables the switch to detect link errors using forward error correction and automatically retransmits packets, while CBFC prevents buffer overflows that traditionally caused packet loss, resulting in a higher level of reliability for data-intensive workloads.

In addition, the Ethernet switch features adaptable and optimized Ethernet headers. This reduces header overhead from 46 bytes down to as low as 10 bytes, while maintaining full Ethernet compliance, Broadcom said, and improves network efficiency and enables application-specific optimizations.

Thanks to these features, the Tomahawk Ultra offers high-efficiency performance for large-scale simulations, scientific computing, and synchronized AI model training and inference.

Ethernet switches have long been perceived as higher-latency and lossy, Broadcom said. The Tomahawk Ultra advances Ethernet for high-performance networking and AI scale-up, setting a new benchmark for tightly synchronized AI compute at scale by enabling sub-400 ns XPU-to-XPU communication latency, including the switch transit time, when deployed with scale-up Ethernet (SUE), the company added.

The Ethernet switch also solves one of the bottlenecks in AI and machine learning workloads through in-Network Collectives instead of using collective operations like AllReduce, Broadcast, or AllGather, Broadcom said. This reduces job completion time and improves utilization of compute resources. This capability also is endpoint-agnostic, so it can be used across a range of system architectures and vendor ecosystems.

Tomahawk Ultra supports advanced HPC topologies including Dragonfly, Mesh, and Torus. It is also compliant with the UEC standard.

Broadcom also introduced SUE-Lite, which is tailored for power and area-sensitive accelerator applications. It retains the key low-latency and lossless characteristics of full SUE, while reducing the silicon footprint and power consumption of Ethernet interfaces on AI XPUs and CPUs. It enables easier integration of standards-compliant Ethernet fabrics in AI platforms.

Shipping now, the Tomahawk Ultra is 100% pin-compatible with Tomahawk 5. Earlier this year, Broadcom released the 102.4 Tbits/s Tomahawk 6 for scale-up and scale-out AI networks.

The post Broadcom delivers Ethernet switch for HPC and AI appeared first on Electronic Products.