“In a backbone and leaf structure, compute assets – racks of servers geared up with CPUs, GPUs, FPGAs, storage, and/or ASICs – are linked to leaf or top-of-rack switches, which then join by way of varied aggregation layers to the backbone,” Google wrote. “Historically, the backbone of this community makes use of Digital Packet Switches (EPS), that are commonplace community switches offered by corporations like Broadcom, Cisco, Marvell, and Nvidia. Nonetheless, these EPS devour a big quantity of energy.”
“Apollo is believed to be the primary large-scale deployment of optical circuit switching (OCS) for knowledge heart networking. The Apollo OCS platform features a homegrown, internally developed OCS, circulators, and customised wavelength-division-multiplexed (WDM) optical transceiver know-how that helps bidirectional hyperlinks by way of the OCS and circulators. Apollo has served because the spine of all Google knowledge heart networks, having been in manufacturing for almost a decade, supporting all knowledge heart use instances.
“Incorporating the Apollo OCS layer replaces the backbone blocks, leading to important value and energy financial savings by eliminating {the electrical} switches and optical interfaces used within the backbone layer. Google makes use of these optical switches in a direct join structure to hyperlink the leaves by way of a patch panel. This technique just isn’t packet switching; it features as an optical cross-connect,” Google said.
“OCS switches provide excessive bandwidth and low community latency, together with a big discount in capital expenditures. This is because of their capacity to cut back the variety of required electrical switches, thereby eliminating expensive optical-to-electrical-to-optical conversions,” mentioned Sameh Boujelbene, vice chairman with the Dell’Oro Group. “Furthermore, in contrast to electrical switches, OCS switches don’t want frequent upgrades when servers undertake next-generation optical transceivers.”
Nonetheless, OCS continues to be an rising know-how. “Thus far, solely Google has managed to deploy them at scale in its knowledge heart networks after a few years of growth. Moreover, OCS switches could necessitate adjustments to the present fiber infrastructure, relying on the cloud service supplier,” Boujelbene mentioned.
“OCS switches have been deployed at Google in backbone layer however with the emergence of AI purposes, we see them being deployed extra contained in the AI clusters due to the advantages that they create,” Boujelbene mentioned.
Standardizing optical transport applied sciences
Necessities for increased pace Ethernet networking tools are evolving as AI networks broaden. For instance, there’s rising demand for 800G Ethernet using 800ZR high-speed optical transmission know-how and OpenZR+, the business initiative to develop interoperable requirements for coherent optical transceivers.
On the 400G Ethernet stage, 400ZR has been “an awesome success for the coherent pluggable business with a number of suppliers and an incredible quantity of 400ZR QSFP-DD and OSFP modules deployed in metro DCI [data center interconnect] purposes,” based on Cisco’s Acacia web site. (Cisco acquired optical maker Acacia Communications for $4.5 billion deal in 2021.)
“Community grade pluggable optics reminiscent of 400ZR and others will see important uptick in deployments in 2024 in communication service supplier networks,” IDC reported lately.
