In the recent “Enabling AI Infrastructure” event organized by Broadcom in mid-March, revolutionary perspectives on strategy and innovation in Artificial Intelligence (AI) infrastructure were presented. As a leader in the Main Switching Group, Broadcom highlighted the crucial role of scalable networks and emphasized why Ethernet technology is the preeminent choice to power AI infrastructure.
Broadcom’s argument revolves around a fundamental question: How to achieve hyper-scale volume effectively? Technological history shows that adopting individual building blocks that interoperate through common standards is the most efficient way to achieve hyper-scale implementation.
Distributed Computing as the Key
AI infrastructure faces the challenge of being essentially a distributed computing problem. Given the capacity limitation of a single GPU, the need for multiple GPUs working together is inevitable. This distributed approach demands a robust network, and this is where Ethernet technology shines, surpassing any other network alternative for its performance, cost-effectiveness, and adoption of open standards.
Ethernet: The Network as the Core of Distributed Computing
For decades, the premise that “the network is the computer,” proclaimed by Sun Microsystems, has proven to be visionary. The network, as the foundation of distributed computing, emerges as the only viable approach from a performance and cost perspective. Google, for example, used Ethernet about 20 years ago to connect hundreds of thousands of CPUs, creating one of the world’s largest distributed computing systems without resorting to centralized computing architectures.
Solution to Network Traffic and Congestion
The challenge of network congestion is monumental, especially when GPUs exchange large volumes of data simultaneously. Broadcom addresses this issue with innovative technologies that optimize traffic management. Among these solutions are:
– The Jericho 3 AI switch: Optimizes clusters of up to 24,000 GPUs, offering 10% better performance at half the cost of existing technologies, ideal for AI network architectures that demand efficient and heterogeneous traffic management.
– The Tomahawk switch chip: Represents the highest bandwidth Ethernet technology, enabling the most powerful and efficient data centers required for the future of AI and hyper-scale cloud infrastructure.
– The Thor 2 Ethernet Network Interface Controller (NIC): Doubles bandwidth capacity to 400 Gigabits per second, integrating advanced end-to-end congestion management capabilities.
Total Connectivity
In conclusion, Ethernet not only provides superior performance to any other alternative network technology but is also the most cost-effective and innovative option, backed by open standards and a vast ecosystem. Broadcom is dedicated to continuously driving the advancement of Ethernet solutions to meet the growing demands of AI infrastructure, supported by a diverse ecosystem and a network that promotes effective interconnection. This collective vision ensures that Ethernet will remain the backbone of AI infrastructure in the foreseeable future.