Based on Broadcom, a single Jericho4 system can scale to 36,000 HyperPorts, every working at 3.2 Tbps, with deep buffering, line-rate MACsec encryption, and RoCE transport over distances larger than 100 kilometers.
HBM powers distributed AI
Bettering on earlier designs, Jericho4’s use of HBM can considerably enhance whole reminiscence capability and scale back the facility consumed by the reminiscence I/O interface, enabling sooner information processing than conventional buffering strategies, in line with Lian Jie Su, chief analyst at Omdia.
Whereas this will likely elevate prices for information heart interconnects, Su stated higher-speed information processing and switch can take away bottlenecks and enhance AI workload distribution, growing utilization of knowledge facilities throughout a number of places.
“Jericho4 could be very completely different from Jericho3,” Su stated. “Jericho4 is designed for long-haul interconnect, whereas Jericho3 focuses on interconnect inside the identical information heart. As enterprises and cloud service suppliers roll out extra AI information facilities throughout completely different places, they want secure interconnects to distribute AI workloads in a extremely versatile and dependable method.”
Others identified that Jericho4, constructed on Taiwan Semiconductor Manufacturing Firm’s (TSMC) 3‑nanometer course of, will increase transistor density to assist extra ports, built-in reminiscence, and larger energy effectivity, options that could be crucial for dealing with giant AI workloads.
“It allows unprecedented scalability, making it supreme for coordinating distributed AI processing throughout expansive GPU farms,” stated Manish Rawat, semiconductor analyst at TechInsights. “Built-in HBM facilitates real-time, localized congestion administration, eradicating the necessity for advanced signaling throughout nodes throughout high-traffic AI operations. Enhanced on-chip encryption ensures safe inter-data heart site visitors with out compromising efficiency.”