Broadcom has launched its latest networking processor, the Tomahawk Ultra, aimed at enhancing artificial intelligence (AI) workloads by improving chip-to-chip communication inside data centres. The launch marks a strategic move by the semiconductor firm to deepen its challenge to Nvidia’s dominance in AI infrastructure.
The Tomahawk Ultra is designed to act as a high-speed traffic controller, managing the vast streams of data exchanged between hundreds of interconnected processors. This function is critical in data centres where “scale-up” computing, linking numerous chips in close proximity, is vital to harnessing the computational power required for advanced AI tasks.
Broadcom’s new chip goes head-to-head with Nvidia’s NVLink Switch, a key component in the GPU-maker’s data centre offerings. However, according to Ram Velaga, Broadcom’s Senior Vice President and General Manager for Core Switching, the Tomahawk Ultra offers a notable advantage: it can connect four times as many chips as Nvidia’s counterpart. Additionally, instead of relying on a proprietary protocol, Broadcom has enhanced standard Ethernet to meet the high-speed demands of AI networking, an approach that could offer broader industry compatibility.
The chip will be manufactured by Taiwan Semiconductor Manufacturing Company (TSMC) using its 5-nanometre process technology. Shipments of the Tomahawk Ultra have already commenced, signalling Broadcom’s readiness to provide a scalable alternative for companies building large-scale AI infrastructure.
Originally designed for high-performance computing (HPC) applications, the processor was repurposed to meet the growing demands of AI developers amidst the rapid rise of generative AI. “As the AI landscape evolved, we realised the architecture we had built lent itself well to scaling the kind of computing modern AI requires,” Velaga told Reuters.
Broadcom already plays a key role in AI chip development through its partnership with Alphabet’s Google, which uses Broadcom’s technology to manufacture its in-house AI accelerators, considered among the few realistic alternatives to Nvidia’s powerful GPUs.
(With inputs from Reuters)
