Within the fast-paced world of community infrastructure, few applied sciences have confirmed as transformative as Phase Routing over IPv6 (SRv6). What began as a method to simplify service supplier networks and help 5G rollouts has now grow to be vital for dealing with as we speak’s most difficult synthetic intelligence (AI) workloads. This thrilling evolution—from overcoming conventional networking challenges to driving cutting-edge AI networks—showcases not solely the exceptional flexibility of SRv6 but additionally its pivotal function in redefining the way forward for community structure. As we embrace this new frontier, SRv6 stands on the forefront, enabling improvements that can form the best way we design AI infrastructures.
The genesis of SRv6: A quest for community simplification
Since 2012, Cisco has been on the forefront of pioneering Phase Routing, serving to pave the best way for SRv6, which started to take form round 2016. This period marked a pivotal second within the business because it acknowledged the pressing want for a extra agile and programmable community infrastructure able to accommodating the calls for of rising applied sciences reminiscent of 5G, Web of Issues (IoT), and cloud providers. The SRv6 community programming mannequin was first launched on the Web Engineering Job Power (IETF) in March 2017, heralding the onset of an ecosystem that has since expanded quickly throughout varied industries.
A key driver behind SRv6 was the aspiration to simplify community operations by harnessing the inherent capabilities of IPv6. In distinction to its predecessor, Phase Routing Multiprotocol Label Switching (SR-MPLS), which nonetheless trusted the MPLS information aircraft, SRv6 sought to function solely inside the IPv6 framework, thereby eliminating the complexities related to multiprotocol environments.
Cisco performed a key function within the early growth of SRv6 by selling its standardization on the IETF. This effort resulted in vital requirements reminiscent of RFC 8402 (Phase Routing Structure), RFC 8754 (Phase Routing Header), and RFC 8986 (SRv6 Community Programming), which established the muse for the expertise. In 2019, Cisco launched the idea of SRv6 uSID (microsegment), enabling large-scale deployments whereas making certain compatibility with older gear.
SRv6 and the 5G revolution
The preliminary driver for SRv6 adoption was clear: The telecommunications business wanted an answer that would meet the stringent necessities of 5G networks. Conventional mobility administration executed via GPRS Tunneling Protocol (GTP) created advanced overlay tunneling architectures that didn’t scale to 5G necessities—elevated numbers of linked gadgets, ultra-low latency calls for, community slicing capabilities, and cellular edge computing. The third Era Partnership Undertaking (3GPP) formally initiated a research merchandise titled “Examine on Person Aircraft Protocol in 5GC” to hunt doable candidates for the subsequent user-plane protocol, with SRv6 rising as a compelling different.
What made SRv6 significantly engaging for 5G was its means to simplify the community stack whereas enhancing capabilities. By leveraging IPv6’s deal with house to supply community programmability, SRv6 enabled operators to compose information paths within the end-to-end IPv6 layer, integrating site visitors engineering, VPNs, and repair chaining options with out the complexity of sustaining per-session tunnel states. Community assets—even wavelengths in dense wavelength division multiplexing (DWDM) techniques—may very well be represented as IPv6 addresses, permitting management planes to program information paths that met particular software necessities.
Fast adoption throughout service supplier networks
Main communications service suppliers (CSPs) have embraced SRv6 and lots of extra are contemplating doing so.
Determine 1: Throughout the globe, a whole bunch of SRv6 initiatives have been deployed or are within the testing or planning phases
These deployments reveal the flexibleness of SRv6 throughout varied purposes:
- Simplified VPN providers: SRv6 makes it simpler to deploy and handle community providers like L3VPNs, even throughout totally different networks. Solely the entry and exit routers must help SRv6, whereas the principle routers can simply ahead commonplace IPv6 site visitors. This streamlines community operations and lowers overhead.
- Service perform chaining (SFC): SRv6 permits community capabilities, like firewalls and cargo balancers, to be included straight in routing paths. This implies you possibly can handle site visitors with out sophisticated further protocols.
- Site visitors engineering (TE) and quick reroute (FRR): SRv6 offers community operators tremendous management over site visitors routes, serving to to satisfy efficiency targets like low latency or bandwidth ensures.
- Operational simplicity and price discount: Through the use of solely the IPv6 framework, SRv6 minimizes the reliance on varied overlay protocols, leading to an easier community. This results in simpler troubleshooting and decrease operational prices.
- Enhanced scalability and aggregation: SRv6 makes use of the scalability of IPv6, making it doable to handle giant networks with fewer prefixes, which simplifies routing and boosts effectivity.
The AI infrastructure problem: A brand new frontier
As SRv6 expertise superior in service supplier networks, a big transformation was additionally going down in information facilities. The speedy progress of AI—and particularly the rise of large-scale mannequin coaching—created networking calls for which are essentially totally different from conventional workloads. AI coaching workloads scale to unbelievable ranges, involving hundreds and even tens of hundreds of graphics processing items (GPUs) working concurrently. Not like conventional information heart site visitors patterns, which encompass numerous and unbiased transactions, AI coaching workloads intensify the long-standing “elephant movement” problem. Whereas elephant flows have existed in large information shuffles, IP storage, and high-performance computing (HPC), AI coaching creates demanding patterns: hundreds of tightly synchronized GPUs executing collective communication operations (all-reduce, all-gather) at each coaching step, producing huge, simultaneous information transfers the place any straggler delays your complete cluster.
This synchronized habits creates vital challenges that conventional networking approaches wrestle to handle:
- Bursty site visitors and congestion spikes: When hundreds of GPUs concurrently push information alongside the identical paths, sudden, intense congestion spikes can happen. Whereas Express Congestion Notification (ECN) stays vital for managing congestion reactively, with out proactive site visitors placement these mechanisms could be overwhelmed, doubtlessly inflicting head-of-line blocking that spreads congestion throughout the community.
- The “slowest packet” drawback: AI community efficiency is dictated by the slowest packet, not averages. When hundreds of GPUs look ahead to a single straggler packet, even slight latency will increase can considerably impression job completion time (JCT). Each microsecond and each dropped packet issues.
- Scale-across complexity: As AI infrastructure extends past particular person information facilities, organizations face community area fragmentation, state scalability challenges at geographic scale, dynamic WAN circumstances, and operational complexity spanning a number of protocol domains.
SRv6 in AI: The pure evolution
The networking neighborhood acknowledged that the identical ideas that made SRv6 profitable in 5G networks—stateless operation, source-driven path management, and unified IPv6-based structure—may deal with AI infrastructure challenges.
Backend GPU cloth optimization employs varied congestion administration methods. Adaptive routing and flowlet load balancing are actively deployed at hyperscalers and neoclouds, offering dynamic site visitors distribution based mostly on real-time community circumstances. SRv6’s uSID presents another strategy via deterministic path placement for distant direct reminiscence entry (RDMA) site visitors. Through the use of a deep integration between AI workloads and SRv6, community interface controllers (NICs) can leverage supply routing to carry out stateless, predictable path placement—explicitly distributing site visitors from totally different sources throughout accessible paths. This deterministic strategy enhances reactive strategies reminiscent of ECN by enabling proactive site visitors placement that may scale back the frequency and severity of congestion occasions. Moreover, SRv6’s specific path encoding simplifies failure restoration: When congestion or failures come up, new paths could be encoded on the supply with out counting on distributed routing convergence, permitting for speedy site visitors movement changes.
Moreover, within the realm of frontend community unification, AI frontend networks should deal with quite a lot of site visitors sorts, together with giant checkpoint writes to distributed storage, telemetry streams, management aircraft messages, and person entry. Every of those site visitors sorts has distinctive efficiency necessities. SRv6 presents a unified framework for implementing high quality of service (QoS), safety insurance policies, and site visitors steering throughout each backend and frontend domains. This streamlining eliminates the complexity related to managing totally different coverage frameworks, permitting for higher effectivity in community administration.
Moreover, SRv6 facilitates scale-across structure enablement by eradicating the normal fragmentation between information heart and WAN domains, which ends up in the creation of unified IPv6-based information planes. Organizations can apply constant insurance policies for managing AI site visitors, whether or not it traverses native materials, frontend networks, or spans huge distances between information facilities. With SRv6, a single phase record can encode paths from supply GPUs via the entire infrastructure to vacation spot GPUs positioned in distant information facilities. Not like Useful resource Reservation Protocol Site visitors Engineering (RSVP-TE) or Multiprotocol Label Switching Site visitors Engineering (MPLS-TE), which rely upon sustaining per-flow state on community gadgets, SRv6 incorporates all routing directions straight inside packet headers. This strategy eliminates state explosion, making it significantly useful for scale-across situations.
Plenty of hyperscalers started innovatively utilizing SRv6 of their AI backend networks to supply fine-grained community path management, maximize community utilization, and ship wonderful cloth resiliency. At Open Supply Summit Europe 2025, Cisco and Microsoft showcased how SRv6 in SONiC permits a variety of information heart use circumstances together with AI backend.
The trail ahead
The journey of SRv6, from its origins in service supplier networks to its promising function in AI infrastructure, illustrates a basic fact: Robust architectural ideas transcend particular use circumstances. The stateless operations, source-driven management, and unified IPv6 framework that simplified 5G networks are the identical ideas that allow deterministic efficiency in AI materials and seamless connectivity throughout geographic boundaries.
As AI continues to broaden—from single-cluster deployments to large-scale architectures spanning continents—the networking challenges will solely develop. Coaching periods that contain a whole bunch of hundreds of GPUs distributed throughout a number of information facilities will demand community infrastructure able to sustaining microsecond-level precision on a worldwide scale.
SRv6’s inherent flexibility and extensibility permit it to adapt to those altering wants. Its programmability permits the introduction of latest community capabilities and site visitors engineering capabilities with out requiring basic architectural modifications. As new AI communication patterns emerge, SRv6 supplies a strong networking basis to help them.
The expertise that simplified 5G cellular networks, enabled community slicing, and streamlined service supplier operations is now the identical expertise making certain that AI infrastructure can scale with out limits. Since its first demonstrations in 2017, SRv6 has confirmed itself not simply as a networking protocol however as a basic constructing block for the way forward for digital infrastructure. As organizations develop the subsequent era of AI techniques, SRv6 will function a strong but unobtrusive engine, serving to make sure that the community stays an enabler of innovation slightly than a bottleneck. The journey from 5G to AI is only the start; the structure is effectively positioned for no matter comes subsequent.
