Broadcom's Jericho4 Chip: Revolutionizing AI Infrastructure with Distributed Data Center Connectivity

Broadcom Unveils Jericho4: A Game-Changer for AI Infrastructure

Broadcom has introduced Jericho4, a groundbreaking networking chip designed to revolutionize artificial intelligence (AI) infrastructure. Unveiled on Monday, this innovative ASIC (Application-Specific Integrated Circuit) aims to connect data centers up to 100 kilometers apart, potentially transforming the landscape of AI model training and deployment 1

1

2

2

.

Key Features and Capabilities

The Jericho4 boasts impressive specifications, including 51.2 Tb/s of aggregate bandwidth across its switch and fabric ports 1

1

. One of its standout features is the introduction of "hyper ports" - four 800GbE links that function as a single 3.2Tb/s port. This configuration allows for up to 70% higher link utilization compared to traditional ECMP link aggregation 1

1

.

Broadcom claims that users can scale Jericho4 into configurations of up to 36,000 hyper ports, enabling a staggering 115.2 petabits per second connection between two data centers. This bandwidth is sufficient to connect 144,000 GPUs at 800Gbps each without encountering bottlenecks 1

1

.

Addressing AI Infrastructure Challenges

The Jericho4 chip arrives at a crucial time when the AI industry faces significant infrastructure challenges. As AI model sizes grow and computational requirements intensify, companies are building massive, power-hungry data center campuses to meet these demands 1

1

3

3

.

Broadcom's solution offers an alternative approach. Instead of constructing enormous single-location facilities, companies could leverage Jericho4 to interconnect multiple smaller data centers, effectively pooling their resources 2

2

. This distributed approach could help alleviate power constraints and provide more flexibility in infrastructure deployment.

Technical Innovations for Enhanced Performance

To address potential latency issues in distributed setups, Jericho4 incorporates several technical innovations:

1. Deep HBM-backed buffers and congestion management technology to mitigate tail latency caused by packet loss 1

   1

   .
2. The use of high-bandwidth memory (HBM), similar to that used in AI processors from Nvidia and AMD, to handle large volumes of data efficiently 3

   3

   5

   5

   .
3. Line-rate, full-speed encryption to ensure data security when transferring information between data centers 4

   4

   .

Broadcom's Vision for AI Networking

Ram Velaga, Senior Vice President and General Manager of Broadcom's Core Switching Group, emphasized the company's commitment to Ethernet as the backbone for next-generation AI infrastructure 4

4

. Broadcom's strategy involves a three-tiered approach to scaling AI infrastructure:

1. Scale-Up Ethernet (SUE) for intra-rack communication
2. Scale-Out Ethernet for larger clusters within a single data center
3. Jericho4 for inter-data center connectivity 4

   4

This approach aims to provide a scalable, vendor-neutral alternative to proprietary AI interconnects, leveraging Ethernet's ubiquity, cost-efficiency, and openness 4

4

.

Industry Impact and Future Prospects

The introduction of Jericho4 could have far-reaching implications for the AI industry. By enabling the connection of geographically dispersed data centers, it opens up new possibilities for distributed AI training and inference 2

2

3

3

. This could lead to more efficient resource utilization and potentially reduce the environmental impact of AI infrastructure.

As AI continues to evolve rapidly, Broadcom's innovation represents a significant step towards more flexible and scalable AI computing solutions. The chip's ability to facilitate large-scale distributed computing may prove crucial as AI model sizes continue to grow and computational demands increase 4

4

5

5

.

With Jericho4 now available for sampling by large customers, the industry eagerly anticipates its potential impact on the future of AI infrastructure design and deployment 1

1

.