Vera Rubin

Type: Product (AI Supercomputer Platform)

NVIDIA's Vera Rubin is a six-chip co-designed AI supercomputer platform that treats the rack — not the GPU — as the fundamental unit of compute. Announced at CES January 2026, with systems shipping H2 2026, it represents a 5x inference improvement and 3.5x training improvement over the previous Blackwell generation.

Platform Specifications

Rubin GPU:

• 50 PFLOPS NVFP4 inference, 35 PFLOPS training
• 336 billion transistors (up from 208B on Blackwell)
• 224 SMs with fifth-generation Tensor Cores
• 288 GB HBM4 at 22 TB/s bandwidth
• NVLink 6 at 3.6 TB/s bidirectional per GPU

Vera CPU:

• 88 custom Olympus cores (Arm), 176 threads via Spatial Multithreading
• 162 MB unified L3 cache
• 1.5 TB LPDDR5X at 1.2 TB/s
• Coherent CPU-GPU link at 1.8 TB/s
• PCIe Gen6 with CXL 3.1

Six Co-Designed Chips:

1. Vera CPU
2. Rubin GPU
3. NVLink 6 Switch (36 per NVL72 rack, SHARP FP8 acceleration)
4. ConnectX-9 (800 Gb/s per port, inline cryptography)
5. BlueField-4 DPU (64-core Grace CPU, 800 Gb/s networking)
6. Spectrum-6 Ethernet (102.4 Tb/s, co-packaged silicon photonics)

NVL72 Rack:

• 72 Rubin GPUs, all-to-all NVLink topology
• 260 TB/s aggregate scale-up bandwidth
• 180-220 kW, fully liquid-cooled
• Cableless modular trays (5-minute assembly)

Performance vs Blackwell

Target Workloads

• Mixture-of-Experts inference (10x lower cost/token vs Blackwell)
• Long-context inference (100K+ tokens)
• Agentic reasoning pipelines
• Continuous training/post-training
• Multi-tenant, multi-model execution

Timeline

• January 2026: Announced at CES, full production confirmed
• March 2026: Vera Rubin POD announced with seventh chip
• H2 2026: Systems shipping to customers

Mentioned In

• Rack-Scale AI Compute — Exemplar of rack-as-product architecture
• HBM4 Memory Architecture — First platform to deploy HBM4
• Silicon Photonics — Spectrum-6 switch integration

Related Entities

• NVIDIA — Parent company