In December 2019, we made a daring announcement about how we’d perpetually change the economics of the web and drive innovation at speeds like nobody had ever seen earlier than. These have been bold claims, and never surprisingly, many individuals took a wait-and-see angle. Since then, we’ve continued to innovate at an more and more quick tempo, main the {industry} with progressive options that meet our clients’ wants.

As we speak, simply three and a half years after launching Cisco Silicon One™, we’re proud to announce our fourth-generation set of units, the Cisco Silicon One G200 and Cisco Silicon One G202, which we’re sampling to clients now. Sometimes, new generations are launched each 18 to 24 months, demonstrating a tempo of innovation that’s two instances quicker than regular silicon improvement.

The Cisco Silicon One G200 affords the advantages of our unified structure and focuses particularly on enhanced Ethernet-based synthetic intelligence/machine studying (AI/ML) and web-scale backbone deployments. The Cisco Silicon One G200 is a 5 nm, 51.2 Tbps, 512 x 112 Gbps serializer-deserializer (SerDes) system. It’s a uniquely programmable, deterministic, low-latency system with superior visibility and management, making it the perfect selection for web-scale networks.

The Cisco Silicon One G202 brings comparable advantages to clients who nonetheless wish to use the 50G SerDes for connecting optics to the change. It’s a 5 nm, 25.6 Tbps, 512 x 56 Gbps SerDes system with the identical traits because the Cisco Silicon One G200 however with half the efficiency.

To realize the imaginative and prescient of Cisco Silicon One, it was crucial for us to spend money on key applied sciences. Seven years in the past, Cisco started investing in our personal high-speed SerDes improvement and realized instantly that as speeds enhance, the {industry} should transfer to analog-to-digital (ADC)-based SerDes. SerDes acts as a elementary constructing block of networking interconnect for high-performance compute and AI deployments. As we speak, we’re happy to announce our next-generation, ultra-high efficiency, and low-power 112 Gbps ADC SerDes able to ultra-long attain channels supporting 4-meter direct-attach cables (DACs), conventional optics, linear drive optics (LDO), and co-packaged optics (CPO), whereas minimizing silicon die space and energy.

The Cisco Silicon One G200 and G202 are uniquely positioned within the {industry} with superior options to optimize real-world efficiency of AI/ML workloads—whereas concurrently driving down the associated fee, energy, and latency of the community with vital improvements.

The Cisco Silicon One G200 is the perfect answer for Ethernet-based AI/ML networks for a number of causes:

~ With the {industry}’s highest radix change, with 512 x 100GE Ethernet ports on one system, clients can construct a 32K 400G GPUs AI/ML cluster with a 2-layer community requiring 50% much less optics, 40% fewer switches, and 33% fewer networking layers—drastically lowering the environmental footprint of the AI/ML cluster. This protects as much as 9 million kWh per 12 months, which in line with the U.S. Environmental Safety Company is equal to greater than 6,000 metric tons of carbon dioxide (CO2e) or burning 7.3 million kilos of coal per 12 months.

~ Superior congestion-aware load balancing strategies allow networks to keep away from conventional congestion occasions.

~ Superior packet-spraying strategies decrease creation of congestion sizzling spots within the community.

~ Superior hardware-based link-failure restoration delivers optimum efficiency throughout large web-scale networks, even within the presence of faults.

Right here’s a more in-depth have a look at a few of our many Cisco Silicon One–associated improvements:

Converged structure

~ Cisco Silicon One offers one structure that may be deployed throughout buyer networks, from routing roles to web-scale front-end networks to web-scale back-end networks, dramatically lowering deployment timelines, whereas concurrently minimizing ongoing operations prices by enabling a converged infrastructure.

~ Utilizing a standard software program improvement equipment (SDK) and customary Swap Abstraction Interface (SAI) layers, clients want solely port the Cisco Silicon One setting to their community working system (NOS) as soon as and make use of that funding throughout numerous community roles.

~ Like all our units, the Cisco Silicon One G200 has a big and totally unified packet-buffer optimizing burst-absorption and throughput in giant web-scale networks. This minimizes head-of-line blocking by absorbing bursts as a substitute of the era of precedence movement management.

Optimization throughout all the worth chain

~ The Cisco Silicon One G200 has as much as two instances greater radix than different options with 512 Ethernet MACs, enabling clients to considerably scale back the associated fee, energy, and latency of community deployments by eradicating layers of their community.

~ With our personal internally developed, next-generation, SerDes expertise, the Cisco Silicon One G200 system is able to driving 43 dB bump-to-bump channels that allow co-packaged optics (CPO), linear pluggable objects (LPO), and the usage of 4-meter 26 AWG copper cables, which is nicely past IEEE requirements for optimum in-rack connectivity.

~ The Silicon One G200 is over two instances extra energy environment friendly with two instances decrease latency in comparison with our already optimized Cisco Silicon One G100 system.

~ The bodily design and structure of the system is constructed with a system-first strategy, permitting clients to run system followers slower, dramatically reducing system energy draw.

Revolutionary load balancing and fault detection

~ Assist for non-correlated, weighted equal-cost multipath (WECMP) and equal-cost multipath (ECMP) load balancing capabilities with near-ideal traits assist to keep away from hash polarization, even throughout large networks.

~ Congestion-aware load balancing for stateful ECMP, movement, and flowlet permits optimum community throughput with optimum flow-completion time and job-completion time (JCT).

~ Congestion-aware stateless packet spraying permits close to very best JCT through the use of all accessible community bandwidth, no matter movement traits.

~ Assist for hardware-based redistribution of packets primarily based on hyperlink failures permits Cisco Silicon One G200 to optimize real-world throughput of large scale networks.

Superior packet processor

~ The Cisco Silicon One G200 makes use of the {industry}’s first totally customized, P4 programmable parallel packet processor able to launching greater than 435 billion lookups per second. It helps superior options like SRv6 Micro-SID (uSID) at full price and is extendable with full run-to-completion processing for much more advanced flows. This distinctive packet processing structure permits flexibility with deterministic low latency and energy.

Deep visibility and analytics

~ Programmable processors allow assist for traditional and rising web-scale in-band telemetry requirements offering industry-leading community visibility.

~ Embedded {hardware} analyzers detect microbursts with pre- and post-event logging of temporal movement data, giving community operators the power to investigate community occasions after the very fact with {hardware} time visibility.

A brand new era of community capabilities

Gone are the times when the {industry} operated in silos. With its one unified structure, Cisco Silicon One erases the laborious dividing traces which have outlined our {industry} for too lengthy. Clients not want to fret about architectural variations rooted in previous creativeness and expertise limitations. As we speak, clients can deploy Cisco Silicon One in a mess of how throughout their networks.

With the Cisco Silicon One G200 and G202 units, we prolong the attain of Cisco Silicon One with optimized high-bandwidth units purpose-built for backbone and AI/ML deployments. Clients can get monetary savings by deploying fewer and extra environment friendly units, take pleasure in new deployment topologies with ultra-long-reach SerDes, enhance their AI/ML job efficiency with progressive load balancing and fault discovery strategies, and enhance community debuggability with superior telemetry and {hardware} analyzers.

For those who’ve been watching since we first introduced Cisco Silicon One in December 2019, it’s simple to see that that is just the start. We’re trying ahead to persevering with to speed up the worth addition for our clients.

Keep tuned for extra thrilling Cisco Silicon One developments.

Be taught extra about

structure, units, and advantages.

Further Assets

Learn my first weblog on Silicon One: Constructing AI/ML Networks with Cisco Silicon One

Share: