AI networking at race speed: Leveraging Multipath Reliable Connection (MRC) to build a coordinated fabric

Artificial intelligence (AI) didn’t break the network. It changed the race.

Traditional Ethernet was designed around independent traffic flows, much like cars traveling on a highway. AI workloads are fundamentally different. Thousands of graphics processing units (GPUs) operate as a synchronized system in which small variations in latency, congestion or packet delivery can impact the performance of an entire computation.

AI networking is therefore less like managing highway traffic and more like controlling a Formula 1 race. Success is not determined by speed alone. It depends on coordination, timing, visibility and the ability to react quickly as conditions change. This is why technologies such as multipath networking, congestion signaling, telemetry, automation and intelligent transport mechanisms are becoming critical components of modern AI fabrics.

The future AI network will not be defined only by how much bandwidth it provides. It will be defined by how intelligently it coordinates traffic across the entire system.

The real constraint is coordination, not bandwidth

A common assumption is that AI networking is primarily a bandwidth challenge. This is not true.

Modern networks already deliver enormous throughput. The greater challenge is to maintain predictable performance when thousands of GPUs generate highly synchronized communication patterns. AI workloads create unique networking behaviors, including:

• Incast congestion, where many senders transmit toward the same destination
• Queue buildup, as synchronized flows compete for shared resources
• Tail latency amplification, where the slowest communication path impacts the entire workload
• Synchronization stalls, where GPUs wait for other participants before continuing

Traditional flow-based mechanisms such as equal-cost multipath (ECMP) remain an essential foundation. However, traffic from high-scale AI training exposes situations where flow-level hashing alone may not fully utilize the available fabric or maintain predictable performance under highly synchronized demand.

This is one of the motivations behind Multipath Reliable Connection (MRC). MRC is an architecture deliver by the Open Compute Project (OCP) that retains RoCEv2 transport while introducing capabilities like packet spraying and source-driven congestion mitigation techniques. By improving the way traffic is distributed and coordinated across multiple network paths, MRC aims to improve fabric utilization and provide more predictable performance for large-scale AI workloads while enabling a practical migration path for existing AI networks.

The challenge is no longer simply moving packets faster. It is to keep all parts of the system moving together. On a Formula 1 track, the problem is not the speed of individual cars. The challenge appears when the entire field reaches the same corner at the same time. Even the fastest cars need coordination to succeed. AI traffic behaves the same way.

Packet spraying: Using the full track

Packet spraying represents an evolution beyond traditional single-path forwarding. Traditional networks often assign a flow to a single ECMP-selected path. This works well for many applications, but traffic from AI GPUs frequently create synchronized traffic patterns where multiple large flows compete for the same resources. Packet spraying distributes packets from a workload across multiple available paths, which increases path diversity beyond traditional flow-based forwarding.

The benefits of packet spraying include:

• Higher fabric utilization through more balanced traffic distribution
• Better handling of synchronized traffic bursts
• Reduced congestion hotspots
• Improved resilience when individual paths experience changing conditions

Returning to the Formula 1 analogy, imagine a that a team of cars represents a flow from a single AI GPU. In a traditional network, every car follows the same racing line. If that line becomes congested, the entire team slows down. Packet spraying allows the team to use multiple racing lines across the track. By utilizing the available space more effectively, the team avoids localized bottlenecks and improves overall performance.

Some implementations extend this concept through dynamic load balancing, incorporating additional path information when making forwarding decisions. However, using every lane on the track is only the first step. Traffic from AI GPUs also require mechanisms that maintain reliability and allow the system to adapt as conditions change.

SRv6: Connecting multiple racetracks

SRv6 is a foundational component of the MRC architecture because it provides the mechanisms for deterministic path steering across multiple network planes. It complements packet spraying by enabling traffic to be forwarded over multiple available paths, while higher-level traffic engineering determines which paths are used.

Using the Formula 1 analogy, packet spraying allows every lane of a racetrack to be used efficiently, while SRv6 connects multiple racetracks into a coordinated racing complex. It provides the routes between the tracks, enabling traffic to traverse multiple network planes as a single fabric.

As AI infrastructure continues to evolve, these same capabilities can naturally extend beyond a single data center to connect distributed AI factories using the same policy-based forwarding model. But today, the immediate value is enabling efficient and deterministic connectivity across multiple network planes within the AI fabric.

Multipath coordination: Turning multiple paths into a system

Using multiple paths creates new opportunities, but it also introduces new challenges. The network must ensure that traffic remains reliable, efficient and adaptable as paths experience changing conditions. This is where another element of MRC becomes important. MRC enhances multipath communication by enabling reliable delivery across multiple paths and allowing transport endpoints to incorporate network feedback when they adapt to changing conditions.

The objective is to enable AI workloads to continue operating efficiently while they experience:

• Congestion
• Packet loss
• Transient network events
• Changing path conditions. 

In Formula 1 terms, packet spraying opens the entire track. MRC provides the coordination that allows every driver to adapt while remaining part of the same race. Rather than finding one perfect path, the goal is to use the entire fabric as a coordinated resource.

Entropy: Keeping traffic distributed

Multipath only delivers its full benefit when traffic remains distributed. This is where the concept of traffic entropy becomes important. Increased traffic entropy creates more disorder, uncertainty, or complexity in traffic and results in higher utilization across the network.

In racing, even the fastest cars naturally converge on the same optimal racing line. When the entire field chooses the same line, congestion appears not because the cars lack capability, but because everyone is competing for the same limited space. AI traffic behaves similarly.

Highly synchronized inter-GPU communication can unintentionally concentrate traffic on a subset of available paths, creating hotspots while other resources remain underutilized. Maintaining effective entropy ensures that traffic continues to take advantage of the entire fabric.

This requires cooperation between transport and network layers through techniques such as:

• Transport-level variation 
• Flow identifiers
• Path diversity mechanisms 
• Multipath forwarding. 

The key principle is simple: Multipath works best when the transport and the network cooperate to maintain distribution.

Coordinated transport: Reacting before congestion occurs

Using multiple paths solves only part of the challenge. Traffic from AI GPUs also require the destination endpoints and the network to react consistently as conditions evolve.

At AI scale, thousands of endpoints and network elements may communicate simultaneously across many paths. Without coordinated feedback, independent reactions can lead to inefficient utilization, unstable traffic behavior and unpredictable application performance.

A coordinated transport model allows endpoints to incorporate network feedback and respond intelligently to changing conditions.

Examples include:

• ECN, which provides early congestion notification before packet loss occurs
• SACK and NACK, which improve loss detection and recovery
• Selective retransmission, which recovers only missing information
• Packet trimming, which improves recovery efficiency during congestion events
• Additional signaling mechanisms, which provide awareness of network conditions

ECN offers one of the clearest examples. Traditional congestion response often begins after packets are dropped. ECN provides an earlier indication of congestion by marking packets when queues begin building, allowing endpoints to adjust before congestion impacts performance. Returning to the Formula 1 analogy: ECN is the yellow flag, not the crash. Drivers receive early warning from the track and adjust before an accident occurs. Similarly, ECN can react proactively rather than waiting until performance has already degraded.

Packet trimming follows the same proactive philosophy. Instead of discarding an entire packet during congestion, the network forwards the useful portion and requires retransmission of only the missing data. This reduces recovery traffic and helps AI workloads maintain throughput under load.
At AI scale, congestion is expected. The objective is not to eliminate it, but to detect it early, communicate it efficiently and recover with minimal disruption.

Intelligence closer to the datapath

As AI networks continue to scale, another emerging industry trend is moving more intelligence closer to the silicon. Modern ASICs are evolving beyond simple packet forwarding. New hardware capabilities enable faster responses to transient conditions such as:

• Temporary link degradation
• Short-lived congestion events
• Latency-sensitive disruptions. 

In Formula 1 terms, this is like moving reaction time from the pit wall directly into the car. The network can respond faster because intelligence is closer to where the event occurs. These capabilities complement, rather than replace, higher-layer transport and control mechanisms.

The direction is clear: AI networking is becoming faster, more adaptive and more distributed.

Visibility: The pit wall advantage

No Formula 1 team wins without telemetry. Every decision made on the track is informed by real-time information from the car. AI networking requires the same principle. Operators need continuous visibility into:

• Queue behavior
• Congestion conditions
• Path utilization
• Workload synchronization
• Overall fabric health. 

Visibility is no longer only about understanding what happened. It is about understanding what is happening now. Nokia has developed an architecture in which streaming telemetry sourced from the Nokia SR Linux network operating system (NOS) and processed by the Event-Driven Automation (EDA) platform turns network visibility into an operational advantage.

Telemetry provides the information. Automation provides the response. Together, they create a closed-loop operating model for AI-scale networking.

The promise of MRC and a new architectural model for the coordinated AI fabric

Nokia believes the next generation of AI fabrics will combine:

• Packet spraying and multipath networking to maximize utilization of available resources
• SRv6 to provide deterministic packet routing across multiple paths
• Transport coordination and feedback mechanisms to improve reliability and adaptation
• Telemetry and automation to provide visibility and closed-loop operations
• Hardware-assisted intelligence to improve response time

AI networking is about how these and other capabilities work together. In Formula 1 terms:

• Packet spraying uses the full track
• SRv6 sprays packets across multiple paths in one coordinated environment
• Multipath coordination keeps every car moving together
• Telemetry provides awareness of changing conditions
• Hardware intelligence improves reaction time

The future AI network will forward traffic faster and coordinate the entire system. Because in AI infrastructures, competitive advantage does not come from one technology alone. It comes from keeping the entire system stable, predictable and synchronized when everything is running at the limit. Please stay tuned as Nokia will have more to say in the area of building AI networks.