Rethinking Ethernet for AI Workloads

Distributed AI training introduces highly synchronized communication patterns, such as all-to-all exchanges and incast during gradient aggregation. These patterns create intense bursts of traffic, uneven link utilization and sensitivity to even minor packet loss or jitter.

Traditional Ethernet handles congestion by dropping packets, relying on retransmission mechanisms at higher layers. While effective for conventional applications, this model breaks down under AI workloads where a single dropped packet can stall multiple GPUs and cause gross GPU underutilization. Addressing this requires a more adaptive and intelligent approach within the network fabric itself.

The Teralynx platform is designed with this shift in mind, embedding AI-aware functionality directly into the switching pipeline. Keysight complements this innovation by providing a realistic, system-level validation environment that ensures these capabilities perform as expected under real-world conditions.

Packet Trimming: A Smarter Approach to Congestion

One of the most impactful capabilities demonstrated is packet trimming. Instead of treating congestion as a binary event that results in packet drops, packet trimming allows the network to respond more tactically. When packet losses happen, the switch can selectively reduce or modify packet payloads while preserving essential metadata, enabling downstream systems to recover or reconstruct the missing information.

This approach fundamentally changes the behavior of the network under stress. Rather than triggering widespread retransmissions that consume bandwidth and increase latency, packet trimming allows the system to maintain forward progress. For AI workloads, this translates directly into improved re-transmission efficiency and reduced job completion times.

Additionally, the Teralynx fabric’s implementation is highly flexible, allowing operators to define policies based on traffic characteristics, congestion thresholds or workload priorities. This adaptability is critical in AI environments, where traffic patterns are dynamic and continuously evolving.

Keysight’s KAI Data Center Builder and AresONE‑M 800GE platform play a central role in validating this functionality. By emulating full-stack AI environments—including NICs, DPUs and distributed workloads—the platform generates realistic traffic patterns such as incast and synchronized bursts. These collective communication patterns enable precise observation of when packet trimming is triggered and how it impacts overall system behavior. This level of validation helps us evaluate how effective packet trimming is for actual AI workloads by being able to see the difference in job completion time.

Auto Load Balancing: Keeping Traffic in Motion

While packet trimming improves retransmission when packets are lost during congestion, Auto Load Balancing focuses on preventing congestion in the first place. During collective operations such as all-reduce, all-to-all, or parameter synchronization, flows are often long-lived, bursty, and highly synchronized across thousands of endpoints. Traditional load balancing based on Equal-Cost Multi-Path (ECMP) hashing struggles in this environment because it relies on static hash functions derived from packet headers (such as 5-tuple fields). This approach assumes a large number of flows are statistically distributed across paths. In contrast, AI traffic often consists of many synchronized flows with similar or identical header fields, which can hash to the same links. This results in uneven link utilization, persistent hotspots, and queue build-up on certain paths, while other links remain underutilized. Additionally, ECMP does not react to real-time congestion conditions, so once a flow is assigned to a congested path, it remains there for its lifetime, exacerbating latency and packet loss.

Auto Load Balancing within the Teralynx fabric introduces a more versatile model. Instead of relying solely on static hashing, the switch continuously monitors real-time network conditions such as per-port queue depth, buffer occupancy, and link utilization. By observing these indicators, the fabric can detect early signs of congestion and imbalance across available paths. It then dynamically adjusts traffic distribution, steering flows or packets away from congested links toward underutilized ones. This can be done at fine granularity and at line rate, enabling rapid reaction to the bursty and transient congestion patterns typical of AI workloads. The result is a more balanced fabric that reduces queue buildup, prevents tail drops, and maintains high throughput even under highly synchronized traffic conditions.

The effectiveness of ALB becomes particularly evident under realistic workload conditions. Using Keysight’s emulation capabilities, the network can be subjected to highly uneven traffic distributions and burst-heavy communication phases. AI workloads have low entropy, with flows that don’t vary too much in the 5-tuple, leading the ECMP hashing algorithm to choose a few links more than others. ALB is free of this setback. With the ability to quickly and effectively adapt, the fabric demonstrates improvements in both throughput and latency consistency.

By operating directly within the switch, ALB eliminates the need for host-based intervention, enabling faster response times and reducing system complexity. This is especially important in large-scale AI clusters, where coordination across thousands of endpoints would otherwise be required.

Ultra Ethernet Transport: A Foundation for AI-aware Networking

Packet trimming and ALB are part of a broader evolution toward AI-optimized Ethernet, embodied in Ultra Ethernet Transport. UET introduces enhancements that make Ethernet more suitable for high-performance, distributed workloads, including improved congestion signaling, reliability mechanisms, and support for collective communication patterns.

Teralynx switches are designed to be UEC-ready, supporting these emerging transport capabilities while maintaining interoperability with existing Ethernet deployments. This ensures that operators can adopt new features incrementally, without disrupting their broader infrastructure.

UET provides the framework within which features like packet trimming and ALB operate cohesively. Together, they enable a network that is both reactive and proactive in managing AI traffic, creating a fabric that delivers more deterministic performance, better resource utilization and improved scalability.

Keysight’s validation extends to the UEC ecosystem as well, generating AI-specific UEC transport patterns and verifying that the network behaves correctly under a wide range of conditions. By benchmarking the entire ecosystem, from switch silicon to endpoints, the collaboration ensures that UET capabilities are fully realized in practice.

From Innovation to Deployment Confidence

Innovation alone is not enough. For hyperscalers and cloud operators, confidence in deployment is equally important. Features like packet trimming and Auto Load Balancing must be validated under realistic conditions to ensure they deliver measurable benefits at scale.

The combination of the Teralynx fabric and Keysight’s KAI Data Center Builder and AresONE‑M 800GE platform provides exactly this capability. It bridges the gap between design and deployment, allowing operators to evaluate performance, identify potential issues and optimize configurations before rolling out infrastructure in production environments.

Ethernet is evolving and, as demonstrated by this collaboration, is fully capable of meeting the increasing demands of AI workloads challenge. Through innovations like packet trimming, auto load balancing, and UET support, the network becomes an intelligent, adaptive component of the AI system.

More importantly, these capabilities are not confined to theory or isolated lab environments. They are validated using realistic workloads, ensuring that they can deliver consistent, reliable performance in real deployments.

For organizations building the next generation of AI infrastructure, this represents a significant step forward. Ethernet, enhanced and rigorously tested, is no longer just a standard—it is a powerful and flexible foundation for AI at scale.

# # #

This blog contains forward-looking statements within the meaning of the federal securities laws that involve risks and uncertainties. Forward-looking statements include, without limitation, any statement that may predict, forecast, indicate or imply future events or achievements. Actual events or results may differ materially from those contemplated in this blog. Forward-looking statements are only predictions and are subject to risks, uncertainties and assumptions that are difficult to predict, including those described in the “Risk Factors” section of our Annual Reports on Form 10-K, Quarterly Reports on Form 10-Q and other documents filed by us from time to time with the SEC. Forward-looking statements speak only as of the date they are made. Readers are cautioned not to put undue reliance on forward-looking statements, and no person assumes any obligation to update or revise any such forward-looking statements, whether as a result of new information, future events or otherwise.