NVIDIA technical article 002
Vera Rubin GPU Platform
How NVIDIA combines compute, scale-up networking, scale-out networking, storage and control into one AI infrastructure platform.01
What it is
Vera Rubin is NVIDIA’s current platform for building large AI systems; it is not a single GPU. The platform connects specialised processors for AI computation, CPU orchestration, rack-scale communication, cluster networking, storage and infrastructure control.
The Rubin GPU is its principal compute engine. Vera CPUs prepare and coordinate work; NVLink 6 joins GPUs inside a scale-up domain; ConnectX-9 and Spectrum-6 connect systems across a scale-out fabric; BlueField-4 handles infrastructure services. NVIDIA added Groq 3 LPX in March 2026 as a seventh processor type for selected low-latency inference paths.
- Scale up
- Join accelerators into one tightly coupled compute domain.
- Scale out
- Join racks or pods through Ethernet or InfiniBand.
- Control plane
- Provision, isolate, secure and monitor infrastructure.
- AI factory
- NVIDIA’s term for an installation that runs AI workloads continuously.
02
Why a platform is needed
At large scale, faster matrix arithmetic alone does not guarantee faster training or inference. Work can stall in memory, CPU preparation, collective communication, cluster congestion, storage or security processing.
| Constraint | Where it appears | Platform response |
|---|---|---|
| Model state | Weights, activations and KV cache exceed one GPU’s memory. | HBM4 on Rubin plus scale-up domains and external context-memory systems. |
| GPU coordination | Collectives and MoE routing repeatedly exchange data among accelerators. | NVLink 6 provides uniform, high-bandwidth communication inside the rack. |
| Host work | Scheduling, tokenisation, tool use and RL environments can leave GPUs idle. | Vera CPUs provide CPU capacity and coherent CPU–GPU connectivity. |
| Cluster traffic | Synchronised GPU flows create short, intense congestion bursts. | ConnectX-9 endpoints coordinate with Spectrum-6 or Quantum-X800 fabrics. |
| Infrastructure work | Networking, storage, isolation and security consume host resources. | BlueField-4 offloads infrastructure services from the compute path. |
03
From processor to POD
04
Seven processor types
| Processor | Principal role | Published technical point |
|---|---|---|
| Rubin GPU | AI and scientific computation. | 336 billion transistors; 288 GB HBM4; 22 TB/s memory bandwidth. |
| Vera CPU | Host execution, orchestration and data movement. | 88 custom Olympus cores with Arm compatibility. |
| NVLink 6 switch | GPU-to-GPU scale-up fabric and collective acceleration. | 3.6 TB/s bidirectional bandwidth per connected GPU. |
| ConnectX-9 | Endpoint for Ethernet or InfiniBand scale-out traffic. | Up to 800 Gb/s per port; 1.6 Tb/s per GPU in NVL72 integration. |
| BlueField-4 | Networking, storage, security and control-plane offload. | 64-core Grace CPU plus ConnectX-9 networking. |
| Spectrum-6 | Ethernet switching across racks and PODs. | 102.4 Tb/s switch silicon with 200 Gb/s SerDes. |
| Groq 3 LPX | Selected low-latency language-inference execution paths. | Added to the platform in March 2026; separate LPX rack-scale system. |
05
How work moves through the platform
- Prepare work.Vera CPU software schedules requests, prepares data and coordinates host-side execution.
- Compute locally.Rubin GPUs read weights and state from HBM4 and execute tensor, attention and scientific operations.
- Exchange inside the rack.NVLink 6 carries activations, expert dispatch and collective traffic among the 72-GPU domain.
- Cross the network endpoint.ConnectX-9 applies RDMA, scheduling, congestion control, isolation and security at the server edge.
- Scale across racks.Spectrum-X Ethernet or Quantum-X800 InfiniBand moves data across the external fabric.
- Offload infrastructure.BlueField-4 processes storage, network and security services independently of application compute.
06
Five rack-scale systems
| System | Function in the POD | What it contains |
|---|---|---|
| Vera Rubin NVL72 | Primary GPU compute and rack-scale communication. | Rubin GPUs, Vera CPUs, HBM4, NVLink 6, ConnectX-9 and BlueField-4. |
| Groq 3 LPX rack | Low-latency processing for selected inference stages. | Groq 3 LPX processors and associated networking. |
| Vera CPU rack | Dense CPU execution for agent tools, simulations and services. | Liquid-cooled Vera CPU compute nodes. |
| BlueField-4 STX | AI-native storage and context-memory access. | BlueField-4 processors, storage interfaces and Spectrum-X connectivity. |
| Spectrum-6 SPX | Rack-to-rack Ethernet or InfiniBand connectivity. | Spectrum-6 or Quantum-X800 switches and optical interfaces. |
POD scale. NVIDIA publishes a 40-rack reference configuration with 1,152 Rubin GPUs, 60 exaflops of compute and 10 PB/s total scale-up bandwidth. These are platform-design figures, not measured application performance.
07
The software layer
The hardware layers are exposed through one programming and operations stack. Software determines model partitioning, communication patterns, resource isolation and achieved utilisation.
- CUDA and CUDA-X
- Programming, compilers and accelerated libraries.
- NCCL and SHARP
- Collective communication and in-network reduction.
- Dynamo and inference frameworks
- Distributed inference scheduling, serving and model execution.
- DOCA and platform management
- Infrastructure, networking, storage, security and lifecycle operations.
08
Terms used in this article
- Endpoint
- The network interface where a server injects or receives fabric traffic.
- Collective
- An operation involving a group of processors, such as all-reduce.
- KV cache
- Attention state retained while a model generates later tokens.
- POD
- A group of specialised racks operated as one larger system.
- RAS
- Reliability, availability and serviceability mechanisms.
- RDMA
- Direct memory transfer across a network with limited CPU handling.
- SerDes
- Serializer/deserializer circuitry used by high-speed links.
- CPO
- Co-packaged optics placed close to switch silicon.
09
Limits of the evidence
- Most architecture and performance data currently comes from NVIDIA.
- “Up to” and generational claims depend on workload, software and configuration.
- The platform evolved from six to seven processor types in March 2026.
- POD figures describe a reference design, not every customer deployment.
- Price and vendor cost-per-token claims are excluded.
10
Primary sources
- NVIDIA Vera Rubin platformOfficial platform overview and system scope.
- Inside the NVIDIA Vera Rubin PlatformProcessors, architecture, data paths and rack integration.
- Inside NVIDIA Rubin GPU ArchitectureRubin compute, HBM4 and memory mechanisms.
- NVIDIA Vera Rubin PODSeven processors, five rack-scale systems and POD reference figures.
- NVIDIA Vera Rubin platform overview videoOfficial system-level visual explanation.