HPE Networking Data Center Presents at Networking Field Day 41

HPE Networking Data Center Evolving AI data Center Architectures with Praful Lalchandani

Watch on YouTube
Watch on Vimeo

AI infrastructure is rapidly evolving from centralized training environments into complex inference-driven systems in which a single workload can traverse front-end, back-end, storage, and security fabrics. HPE Networking’s strategy is to deliver open, secure Ethernet for every AI fabric (scale-up, scale-out, and scale-across), resulting in high performance, resilience, operational consistency, and strong customer momentum. Praful Lalchandani, Vice President of Product Management for Data Center Networking at HPE Networking, opened the company’s Networking Field Day 41 presentation by citing deployments at a tier-one model provider’s million-GPU cluster, Oracle, neocloud Vultr, and Oak Ridge National Laboratory’s next supercomputer, Lux.

Lalchandani argued that HPE Networking is the only OEM with a product for every AI fabric. Its lineup includes switch trays for AMD’s Helios rack-scale scale-up design, which will ship within months, the QFX family for scale-out, storage, and front-end networks, MX and PTX platforms for data center interconnect and scale-across, and SRX for security. He tied that breadth to rising GPU bandwidth demands. Hopper-class systems needed 400 Gbps per GPU in 2024, Blackwell B300 needed 800 Gbps by the end of 2025, and the next step is 1.6 Tbps roughly 16 months later, with rack-scale designs reaching 14.4 Tbps per GPU. He also highlighted the shift from single-connection rail-optimized designs to multiplane fabrics with two to eight planes. These fabrics enable larger two-tier clusters and let jobs keep running at reduced capacity after a link failure instead of rolling back to a checkpoint.

Lalchandani identified inference, more than training, as the main driver of architectural change, pointing to mixture-of-experts models, chatty agentic workloads, multimodal traffic, disaggregated prefill and decode, and remote KV caching. He acknowledged that nanosecond-scale switch latency is minor next to WAN and GPU time, but said congestion stalls still hurt inter-token latency and tokens-per-second throughput. He traced a single query through the front end, a RAG vector database, prefill nodes, a KV cache store, and decode nodes. That walkthrough showed how caching turns a compute problem into a high-performance data access problem on the storage network. The result is five fabrics with distinct requirements, from multi-tenant isolation at the front end to ZR optics and encryption for scale-across, and each is a point where a stall can degrade the user experience.

Personnel: Praful Lalchandani

HPE Networking Data Center Scale-up Networking with Ethernet for Helios

Watch on YouTube
Watch on Vimeo

AMD Helios delivers massive, rack-scale AI architectures, backed by HPE expertise in liquid cooling, system integration, and high-performance networking. Helios effectively announces the arrival of Ethernet into the scale-up domain, which has been dominated by proprietary protocols. Customers will now benefit from the Ethernet ecosystem that relentlessly drives down costs, pushes innovation forward faster, and brings operational consistency across all AI fabric domains. At Networking Field Day 41, Mike Wong, senior director of product management for Helios rack networking at HPE Networking, detailed the rack and the scale-up switch tray HPE built for it.

Wong described Helios as an open-standard alternative to single-vendor rack-scale systems. It is built on the double-width Open Rack Wide specification and holds 72 AMD MI455 GPUs with 432 GB of HBM each, or 31 TB per rack. The scale-up network uses UALink over Ethernet, so GPUs keep memory-access semantics while running on the Ethernet ecosystem. Each GPU gets 3.6 TB/s of bidirectional bandwidth, and the rack totals 260 TB/s. Wong confirmed for a delegate that these figures are in bytes, not bits. AMD Vulcano NICs provide up to three 800 Gbps scale-out connections per GPU, and Pensando Salina handles front-end and storage traffic. HPE also supplies the management network. The switch tray occupies a single rack unit. It pairs two 102.4 Tbps Tomahawk 6 chips for 204.8 Tbps of raw bandwidth, uses direct liquid cooling on the silicon and retimers, and connects through 200G flyover cables to deliver full-mesh, non-blocking, single-hop GPU connectivity.

Wong said HPE drew on its experience delivering the industry’s first liquid-cooled Tomahawk 6 switch to make around a thousand power and signal integrity changes to AMD’s reference design. He framed this as stability work, because network disruptions waste expensive GPU time. HPE also plans to extend its Mist AI operations tooling from scale-out into the rack, giving operators end-to-end visibility without blind spots. HPE will assemble, test, and crate racks for worldwide drop shipment and provide deployment services that cover cooling distribution units. Wong added that a single rack isn’t meaningful on its own. Customers need a scale-out design suited to their cluster size and load balancing, which HPE supports through Juniper Validated Designs.

Personnel: Mike Wong

HPE Networking Data Center Exploring Evolving Congestion Management in Scale-Out Networks

Watch on YouTube
Watch on Vimeo

Managing congestion in AI training and inference fabrics is critical to maximize GPU utilization, minimize latency and jitter, and protect model accuracy and throughput. Ultimately, congestion-free networks are required to deliver the best end user experiences in the era of agentic AI. This session examines the evolution from dynamic and global load balancing to RDMA-aware load balancing, intelligent packet spraying, and emerging Ultra Ethernet mechanisms. At Networking Field Day 41, Praful Lalchandani, Vice President of Product Management for Data Center Networking at HPE Networking, led the session.

Lalchandani explained that AI fabrics carry a small number of very large flows, so static ECMP hashing leaves some links idle while others drop packets. He split the solution into congestion avoidance through better load balancing and congestion control through fast signaling. HPE’s pre-UEC avoidance work began with dynamic load balancing, which chooses paths based on link utilization and queue depth in hardware. Global load balancing extended this to end-to-end path quality, much like a navigation app. HPE then introduced RDMA-aware load balancing, which splits an RDMA transfer into subflows identified by queue pair and pins each one to a specific spine using BGP link coloring, an approach HPE published through the IETF. In testing, RLB came close to the theoretical maximum of about 370 GB/s with consistent results across runs and eliminated ECN, PFC, and CNP events. If a link fails, only the flows on that link fall back to dynamic load balancing.

Lalchandani said HPE has tested Ultra Ethernet Transport on its QFX switches but has not yet seen production-grade UET implementations on NICs. HPE has, however, adopted UEC concepts such as packet trimming and selective acknowledgment for RDMA traffic. He sees the market moving toward MRC, which borrows UEC’s congestion techniques but keeps the standard RDMA frame format. MRC takes advantage of multiplane fabrics through ECMP-based spraying or SRv6 path pinning, an approach he likened to RLB. Tests with AMD MI350 GPUs, Pollara NICs, and 800G QFX switches put MRC ahead of DLB and ECMP, while SRv6 testing is still underway. He challenged vendor claims of 1.6x or 1.8x Ethernet performance gains, arguing that those figures compare against static ECMP. In his view, open-standard Ethernet already approaches theoretical maximums, and HPE publishes its benchmark results in public validated designs.

Personnel: Praful Lalchandani

Tackling the Power Challenge in AI Data Centers with HPE Networking

Watch on YouTube
Watch on Vimeo

With power becoming a hard constraint on AI expansion, every watt consumed by the network reduces the power available for productive GPU compute. Learn how efficient design in silicon, optics, liquid cooling, and visibility & analytics minimize power consumption. At Networking Field Day 41, Yogesh Kumar, Senior Director of Product Management for Data Center Networking at HPE, framed network efficiency as a matter of compute capacity rather than sustainability. Facility power can take years to provision and is capped once delivered, so every watt the network saves can go back to GPUs.

Kumar described four levers. The first is staying on current silicon generations, since each one improves power per bit. The second is optics: moving from fully retimed modules to linear receive and linear pluggable optics, and eventually to near-packaged and co-packaged designs. The third is liquid cooling, which eliminates fans and improves power usage efficiency. The fourth is network visibility. He highlighted the emerging XPO form factor, which bundles the equivalent of eight OSFPs into a liquid-cooled package. HPE views CPO and NPO as close variants and has experimented with them since the Tomahawk 5 generation, but it has not announced a CPO product. Delegates pressed on serviceability, since CPO lacks field-replaceable parts. Kumar noted that CPO links appear more reliable and that lasers can stay pluggable, while acknowledging that scale-up may eventually require co-packaged approaches.

The cooling discussion centered on the QFX5250, a fanless 102.4 Tbps switch with 64 ports of 1.6T in an ORv3 form factor, which began shipping in May ahead of other OEMs. Kumar said reaching 95% liquid cooling was straightforward, but the last 5% required heat pipes and copper paths to cool components that air would normally cool incidentally. That extra work spares customers rear-door heat exchangers. He credited HPE’s experience with coolant selection, supplier testing, and leak detection, which helped the team avoid over-engineering. For example, the team replaced sensor wiring along every pipe with drip plates and sensors placed at the joints where leaks actually occur. Kumar closed with HPE Data Center Assurance’s network sustainability analytics, which reports power and carbon data, forecasts consumption, and recommends actions such as shutting down unneeded ports, with the stated benefits and risks of each action.

Personnel: Yogesh Kumar

HPE Networking Data Center Scale Across with PTX & Customer On Ramp with MX

Watch on YouTube
Watch on Vimeo

The growth of distributed inference, AI agents, and multisite AI factories is transforming the WAN from basic connectivity into a critical extension of the AI infrastructure. Learn how HPE Juniper Networking’s PTX platforms provide high-capacity, efficient interconnection across AI clusters and sites, while the programmable MX portfolio gives enterprises and service providers a scalable, secure on-ramp connecting customers, applications, and edge environments to AI resources. At Networking Field Day 41, Pradeep Chalicheemala, who leads PTX product management in HPE’s routing business unit, took the discussion beyond the data center.

Chalicheemala used a typical neocloud network to show how data moves through the WAN. Data first arrives from enterprises, telco and edge sites, and autonomous vehicles, and he noted that about 30% of enterprise data is already AI-generated. It is then replicated across geographically dispersed clusters, which also exchange model checkpoints during distributed training. Increasingly, training jobs span sites, and HPE ran an internal training job across clusters 500 kilometers apart. He listed the WAN requirements as low power and cost per bit, converged optical routing, massive throughput, full encryption, and deep routing expertise across heterogeneous networks. The spotlight product was the PTX12K, built on fifth-generation Express silicon. It offers 648 ports of 800GbE with deep buffers, coherent optics on every port, line-rate MACsec, air cooling, and support for 1.6T ports. With a passive mux, 32 ports can share a single fiber across thousands of kilometers without separate transponders.

For customer on-ramp, Chalicheemala presented the MX portfolio, built on programmable Trio silicon now in its sixth generation, with products underway on the seventh. Programmability lets engineers add support for new protocols after a router ships. MX’s large forwarding tables and thousands of VRFs let a single device host thousands of isolated customers, each with its own billing and quality-of-service levels, and its tunneling options carry traffic from the edge deep into the data center. He said MX has served as the on-ramp at most hyperscalers for a decade, and HPE is investing in both silicon lines with roadmaps covering the next three to five years.

Personnel: Pradeep Chalicheemala

HPE Networking Data Center AI Data Center Operations in Action

Watch on YouTube
Watch on Vimeo

As AI clusters grow in size and importance, seemingly minor network issues can cause costly job slowdowns or failures and become increasingly difficult to diagnose through conventional operational methods. Through customer examples and demonstrations, this session shows how HPE Networking Data Center Director and Mist Data Center Assurance combine intent-based automation and agentic AI to simplify deployments and reduce operational complexity across AI fabrics. At Networking Field Day 41, Sridhar Narahari, product manager at HPE, led the session.

Narahari began with a demo of HPE’s AI assistant troubleshooting a job reported as running slowly. The assistant drew on switch telemetry from Apstra Data Center Director, an agent running on the GPU NICs, and job data from schedulers such as Slurm. Within seconds, it traced the bottleneck to the GPU NIC, citing congestion notifications as evidence and confirming that the switches showed no congestion. Narahari attributed the approach to the Mist AI platform HPE is extending into the data center, a context graph that keeps responses grounded, specialized models, and live diagnostics that can run CLI commands. For deployments, he cited Oak Ridge National Laboratory, which built a cluster of more than 4,000 GPUs and 8,000 links in four to five weeks despite having no prior experience with the platform. Operators define intent, such as GPU counts, link speeds, and three-stage or rail-optimized designs. The system then generates what can be hundreds of thousands of configuration lines, validates them in a digital twin before commit, and continuously checks cabling against the design using LLDP data.

Narahari argued that at scale, job failures are a question of when rather than if. He cited Meta data showing a 10-GPU job fails on average after about 48 days, while a job of roughly 1,000 GPUs fails in about eight hours, with optics and links accounting for a growing share of those failures. HPE’s predictive analytics flags optics likely to fail, reporting confidence, time to failure, and impact. It also explains the root cause in plain language, such as declining transmit power on a single lane, and suggests actions such as cleaning a dirty transceiver. Asked about Juniper TAC integration, Narahari said the assistant can open tickets once the operator approves. A colleague added that the model has been trained on past TAC tickets and documentation and can run self-diagnostics on the switches.

Personnel: Sridhar Narahari

Bringing it All Together with HPE AI Data Center Solutions

Watch on YouTube
Watch on Vimeo

An AI factory is more than a collection of GPUs: it’s a scalable, multi-tenant system that integrates infrastructure, operating environments, control-plane services, AI workloads, security, lifecycle management, and observability. This session brings the full HPE AI data center architecture together, illustrating how HPE is the only company that offers integrated compute, storage, and networking and validated, full-stack AI data center solutions. At Networking Field Day 41, Praphul Menon, director of solutions product management for HPE’s AI factory and AI data center solutions, presented the architecture, and Praful Lalchandani closed the session.

Menon said customer conversations over the past two years have made multi-tenancy the defining requirement, with security and performance isolation close behind, because clusters now serve multiple workloads and users dynamically. HPE’s solution has four layers. Infrastructure and an operating layer of bare metal, hypervisors, or Kubernetes form the base. A control plane on top provides management, services, and orchestration, and the top layer runs AI workloads ranging from NVIDIA NIM and Hugging Face models to in-house and Slurm-based code. Security, lifecycle operations, and observability span all four layers. On the networking side, QFX switches and MX and PTX routers provide the fabric, Data Center Director orchestrates networks, and SRX firewalls can isolate tenants from one another. Engagements start with templates that turn GPU counts and rack density into a full bill of materials and topology, down to transceivers and cables. In a demo, Menon created a tenant. Ansible AWX ran playbooks that pulled network segments from NetBox, used Data Center Director APIs to build access and storage networks and routing zones, and assigned them to the tenant’s server.

Under delegate questioning, Menon explained that full tenant creation also provisions operating systems through tools such as Base Command Manager or HPCM and creates isolated storage on Weka or VAST backends. Morpheus provides the tenant administrator’s experience, while operators keep access to the individual element managers. Because the playbooks are atomic, customers can add steps such as approvals or swap NetBox for ServiceNow or their own IPAM system, and the whole stack is sold as one solution. Lalchandani closed by positioning HPE Networking as a best-of-breed pure-play vendor that can now also bring HPE compute, storage, optics, rack integration, and delivery services. He cited Oak Ridge National Laboratory as a turnkey deployment that included turn-up and benchmarking. The result, he argued, is faster GPU productivity, lower integration risk, and a single point of accountability when problems arise.

Personnel: Praful Lalchandani, Praphul​ Menon