Particle.news

Download on the App Store

Broadcom Ships Jericho4 Chip to Accelerate Distributed AI Clusters

Early partner integrations demonstrate its ability to link massive, geographically dispersed GPU clusters with encrypted, congestion-free Ethernet at scale.

The Broadcom company logo is shown outside one of their office complexes in Irvine, California, U.S., March 4, 2021.  REUTERS/Mike Blake/File Photo
Image

Overview

  • Broadcom has begun shipping its next-generation Jericho4 networking ASIC and partners including Nokia and Accton have integrated it into high-performance routers to connect AI workloads across regional data centers.
  • Jericho4 delivers up to 51.2 Tbps of lossless Ethernet fabric and uses high-bandwidth memory buffers to smooth traffic and support interconnection of over one million specialized processing units across multiple sites.
  • The chip secures inter-site data transfers with line-rate MACsec encryption on every port, ensuring performance is maintained while protecting sensitive AI traffic from interception.
  • Fabricated on TSMC’s 3-nanometer process, Jericho4 offers roughly 40% power reduction per bit to help contain energy costs in sprawling, power-hungry AI infrastructures.
  • A single Jericho4 system can scale to 36,000 HyperPorts, consolidating four 800 GbE links into a 3.2 Tbps logical port and enabling interconnect bandwidths that can exceed 115 petabits per second.