As AI advances, networks need to be configured with architectures that support increasingly complex cluster-scale workloads. At Vultr, we’re continuing to collaborate with best-in-class industry partners whose services complement our high-performance global cloud infrastructure – including Broadcom for optimal network performance.
Vultr and Broadcom’s partnership, building on our recent introduction of HPE Juniper Networking to the Vultr Cloud Alliance, enables AI cluster-scale workloads to run over a lossless, congestion-controlled fabric with sustained full line-rate throughput and maximum efficiency. With advanced scheduling, shared-buffer behavior, and RDMA efficiency, Broadcom Tomahawk 5 silicon and Thor 2 NICs (Network Interface Cards) boost network throughput, determinism, and congestion stability.
The switching silicon and NIC behavior help sustain line-rate performance and prevent congestion collapse in AI backend networks, meaning no GPU idle time, more predictable performance, and faster time-to-production.
Broadcom Tomahawk 5 and Thor 2: Designed for consistency amid heavy AI traffic
AI clusters generate extremely high-throughput, continuous east-west traffic synchronized across many GPUs simultaneously. Not only do these workflows need to process data via RoCEv2 (RDMA over Converged Ethernet) fast enough to keep pace with GPUs’ processing capabilities – achieved in part by HPE Juniper Networking Ethernet fabric – but they also need high bandwidth from those GPUs in the first place.
That’s where Broadcom’s joint Tomahawk 5 silicon and Thor 2 NIC solution steps in. By combining Tomahawk 5’s advanced load-balancing capabilities with Thor 2’s RoCEv2 hardware offload, ultra-low latency, and hardware-based congestion control, the solution eliminates congestion and delivers sustained full line-rate throughput for cluster-scale AI workloads.
The Virtual Output Queue (VOQ) scheduler in Tomahawk 5 allocates high bandwidth to each GPU, including amid heavy traffic, a core aspect of AI workloads. Its shared buffer architecture provides access to the full packet buffer across all ports, eliminating packet drops during sustained high loads. Thor 2, with RDMA-enabled GPU-to-GPU direct support, provides sustained 400G, high-volume data transfers. This combination of fast data transfer, data retention, and scheduling maximizes GPU utilization, helping avoid wasted compute power and spend.
Tomahawk 5 uses an extremely high-density 5nm fabrication process, along with high port integration and the option for co-packaged optics (CPO). Also built on a 5nm process, Thor 2 is the most power-efficient 400G NIC, with up to 30% lower power consumption.
Together, Tomahawk 5 and Thor 2 provide a cohesive, field-proven solution with maximum performance, broad ecosystem compatibility, and superior interoperability at lower cost.
Broadcom + HPE Juniper Networking on Vultr infrastructure for maximum AI throughput
Broadcom Tomahawk 5, which powers HPE Juniper Networking’s QFX5240 switch and Thor 2, support backend, frontend, storage, and OOB networks, especially in rail-optimized backend fabrics for AI workloads. Broadcom Tomahawk 5 and Thor 2 reduce congestion and sustain full line-rate data flows through:
- Advanced load balancing
- VOQ scheduling
- Shared buffers
- High-density 800G deterministic switching
Vultr is the cloud infrastructure layer for operationalizing that joint configuration, with on-demand access to cloud GPUs, cloud compute, bare metal servers, and managed Kubernetes, with predictable pricing and easy scalability.
Together, Vultr, HPE Juniper Networking, and Broadcom provide higher GPU utilization, more predictable training time, and stable throughput at AI cluster scale. Deployments are repeatable through Vultr’s simple control panel or API. Thor 2 NICs reduce the need for CPU overhead. And Juniper Validated Designs (JVDs) provide proven, rigorously vetted architectures from AI/ML lab interoperability testing, accompanied by detailed configuration guidance, so customers can know their AI workloads will continue to perform reliably and consistently.
Get started today
Curious how Broadcom and HPE Juniper Networking can elevate your AI deployments on Vultr? Learn more about our partnership, or contact the Vultr sales team.