First 51.2-Tbit/s co-packaged optical Ethernet switch for scalable AI
Broadcom is shipping the industry’s first 51.2 terabits per sec (Tbit/s) co-packaged optics (CPO) Ethernet switch to customers for scalable AI systems.
Bailly integrates eight silicon photonics 6.4-Tbit/s optical engines with Broadcom’s StrataXGS Tomahawk5 switch chip and operates at 70% lower power consumption with an 8x improvement in silicon area efficiency over pluggable transceivers.
Bailly integrates hundreds of optical components and hundreds of millions of transistors in a single optical engine. The high degree of integration enables the placement of the optical engines on a common substrate with complex logic ASICs minimizing the need for signal conditioning circuitry.
The 4RU system design, first shown this time last year, with high-efficiency air cooling provides 128 ports of 400G FR4 connectivity externally fibre-coupled with 128 duplex LC optical connectors with a CPO engine to front-panel routing that supports traditional single-mode fibre.
Broadcom supplies most data centre switch makers (see comments below) and so has a system design compatible to support multiple remote laser modules (RLM) for field replaceability. It has spent the last year reducing the power consumption from the 30% reduction shown in a prototype system last year.
The optical interconnect is critical for both front-end and back-end networks in large scale generative AI clusters. Today, pluggable optical transceivers consume approximately 50% of system power and constitute more than 50% of the cost of a traditional switch system.
The growing bandwidth requirement for the latest GPUs from Nvidia and AMD, coupled with the ever- increasing sizes of AI clusters, requires significantly more power-efficient and cost-efficient optical interconnects that extend beyond discrete implementations.
The integration in the co-packaged optics and silicon photonics platform reduces the latency and boosts the bandwidth while reducing the power, which also helps with the reliability and lifetime.
Broadcom also uses advanced packaging technologies and a highly automated high-density, edge-coupled fibre attach capability.
The company is co-designing platforms with cloud service providers (CSPs) and system integrators to accelerate adoption of CPO platforms and will show the Bailly 51.2T CPO system at the Optical Fiber Communication (OFC) 2024 exhibition at the end of the month.
“As AI clusters demand higher bandwidth density, lower power consumption and lower latency, we are pleased to announce delivery of the industry’s first 51.2-Tbps CPO switch,” said Near Margalit, Ph. D., vice president and general manager of the Optical Systems Division at Broadcom. “Bailly will enable hyperscalers to deploy lower-power, cost-efficient, large-scale AI and compute clusters. Broadcom’s technology leadership and manufacturing innovations help Bailly deliver 70% better power efficiency and ensure an optical I/O roadmap that can walk in tandem with the future bandwidth and power needs of AI infrastructure.”
“Pluggable optics are expected to account for an increasingly significant portion of power consumption at a system level, exceeding 50% of the switch system power at 51.2 Tbps and beyond. This issue will be further exacerbated as Cloud Service Providers build their next- generation AI networks and continue to push for higher speeds,” said Sameh Boujelbene, vice president at Dell’Oro Group.
“Substantial investments in AI infrastructure are accelerating the development of innovative optical connectivity solutions such as Broadcom’s Bailly co-packaged optics platform, aiming to meet the demands of AI clusters while solving cost and power consumption hurdles.”
“Innovation in the optical interconnect will be of great importance to future generations of AI networks,” said Feng Luo, head of network systems at ByteDance. “Broadcom’s continued progress in the development of co-packaged optics is exactly what the industry will require to overcome cost and power bottlenecks of current generation optical interconnects.”
“System design has become more challenging with each increase in SerDes speed, but only recently have we had to re-design our systems and optics to overcome the physical limitations of copper,” said Richard Li, general manager of H3C’s switch product line. “Our partnership with Broadcom to design systems with highly integrated co-packaged optics is a huge step forward to overcoming the design limitations of pluggable transceivers in high density systems.”
“Artificial Intelligence (AI) has introduced a set of challenges that the networking must address,” said Mansour Karam, global vice president (GVP) of data centre products at Juniper Networks. “We are excited to partner with Broadcom to deliver co-packaged optics throughout the data centre, thereby improving cost, power consumption and bandwidth efficiency of the optical interconnect.”
“System design has become more challenging with the need to meet the requirement of AI/ML and high-performance large-scale datacentre networks,” said Grant Lai, chief technology officer (CTO) of Micas Networks. “Our partnership with Broadcom to design systems with highly integrated co-packaged optics will enable a more power efficient future network. The 51.2 Tbps switch that Micas co-developed with Broadcom will reduce communication latency and power consumption in data centres and unleash more AI power.”
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
