Vera-Rubin NVL72 Platform Overview#

Vera-Rubin NVL72 is NVIDIA’s rack-scale Vera CPU and Rubin GPU platform for large AI training, inference, and HPC workloads. The system preserves the NVL72 rack-scale programming model while updating the CPU, GPU, NVLink, networking, power, and cooling architecture.

Use the Vera Rubin NVL72 product specifications for published platform capabilities. Actual memory, storage, link configuration, and enabled features depend on the qualified system configuration.

Rack-Scale Architecture#

A full Vera-Rubin NVL72 rack contains:

  • 72 Rubin GPU packages.

  • 36 Vera CPUs.

  • 18 compute trays.

  • 9 NVLink 6 switch trays.

  • 36 NVLink 6 switch ASICs.

  • 18 north-south BlueField-4 DPUs.

  • 144 ConnectX-9 east-west NIC ASICs.

  • A passive rack backplane implemented as four serviceable cable cartridges for rack-scale NVLink connectivity.

Each compute tray contains two Vera CPUs and four Rubin GPUs. The compute tray is built from two host processor motherboards (HPMs), where each HPM contains one Vera CPU and two Rubin GPUs connected through NVLink-C2C. The tray also contains four ConnectX-9 boards with eight total CX9 NIC ASICs for east-west scale-out networking, a BlueField-4 DPU with dual 400 Gb/s ports for north-south connectivity, four Gen6 E1.S data SSDs, one Gen5 E1.S boot SSD, baseboard management, and tray-level system management.

Each switch tray contains four NVLink 6 switch ASICs. Nine switch trays combine with the passive rack backplane to form a single 72-GPU L1 NVLink domain. For application placement, this means intra-rack GPU collectives should be treated as NVLink-domain traffic first and network traffic second.

Vera CPU#

A Vera CPU socket has 88 NVIDIA Olympus Arm-compatible cores and supports two hardware threads per core using Spatial Multithreading (SMT2). With both threads enabled, a two-socket compute tray exposes 176 physical cores and 352 logical CPUs to the single tray operating system. SOCAMM LPDDR5X capacity is configuration-dependent; record installed and OS-visible capacity rather than assuming a CPU-family maximum is installed in every tray. Vera provides 96 Gen6-capable I/O lanes per socket in the Vera-Rubin configuration before platform-specific routing.

The Vera CPU complex introduces higher core count, SOCAMM LPDDR5X memory, Scalable Coherency Fabric behavior, PCIe Gen6 I/O, NVLink-C2C to Rubin GPUs, and coherent GPU memory modes. Performance-sensitive applications should review CPU affinity, page size, AutoNUMA, IOMMU, compiler tuning, and power controls before drawing platform conclusions.

Use Cases And Workloads#

Large AI workloads continue to scale through pre-training, post-training, and test-time compute. Vera-Rubin NVL72 is designed for those rack-scale patterns: large collectives, dense model parallel communication, high-throughput inference serving, and CPU-assisted data movement near the GPU fabric.

_images/ai-scaling-laws.png

Figure 1 AI scaling laws continue to drive rack-scale GPU system design.#

When moving software to Vera-Rubin NVL72, retune at the rack boundary. CPU core count, memory behavior, PCIe generation, NVLink generation, NIC generation, DPU generation, power envelopes, and liquid-cooling constraints all changed.