Vera Rubin rack

A Vera Rubin rack is an advanced, rack-scale server system built by NVIDIA to power large-scale agentic AI and complex reasoning workloads.

• Memory now accounts for a much larger share of the cost of Nvidia’s newest AI systems. According to Morgan Stanley estimates, it could cost roughly $2 million in a $7.8 million Vera Rubin rack, or about a quarter of the total.

• Vera Rubin (1928–2016) was a pioneering American astronomer whose groundbreaking research on galaxy rotation curves provided the first convincing evidence for the existence of dark matter.

Key Features of Vera Rubin Racks

Agentic AI Design: Built to handle multi-step problem solving and long-context workflows.
Cable-Free Build: The compute trays slide in with no loose cables, hoses, or fans, connecting through a printed circuit board midplane.
Liquid Cooling: Fully liquid-cooled to handle extreme compute densities efficiently.
Automated Assembly: The hardware trays are manufactured using automated robotics for rapid deployment.

The Five Rack Types

The Vera Rubin platform combines five purpose-built rack systems to act as a single supercomputer:

NVIDIA Vera Rubin NVL72: The primary compute rack integrating Vera CPUs and Rubin GPUs.

NVIDIA Vera CPU Rack: Provides dense, liquid-cooled CPU capacity for code execution and tool use.

NVIDIA Groq 3 LPX: Optimized for high-throughput decode inference.

NVIDIA Vera BlueField-4 STX: Manages storage operations and infrastructure security.

NVIDIA Spectrum-6 SPX Ethernet: Handles high-speed, scale-out network connectivity.

Generational Comparison: Blackwell Ultra vs. Vera Rubin NVL72

Specification	GB300 NVL72 (Blackwell Ultra)	VR NVL72 (Vera Rubin)
GPU Count	72 Blackwell Ultra GPUs	72 Rubin GPUs
CPU Count	36 Grace CPUs	36 Vera CPUs
CPU Cores	72 ARM cores per CPU	88 Olympus ARM cores per CPU
FP4 Inference Performance	1.44 ExaFLOPS	3.6 ExaFLOPS
NVFP4 per GPU (Inference)	20 PFLOPS	50 PFLOPS
NVFP4 per GPU (Training)	10 PFLOPS	35 PFLOPS
GPU Memory Type	HBM3e	HBM4
GPU Memory Bandwidth	~8 TB/s	~22 TB/s
NVLink Generation	NVLink 5	NVLink 6
NVLink Bandwidth (per GPU)	1.8 TB/s	3.6 TB/s
Rack-Scale NVLink Bandwidth	130 TB/s	260 TB/s
Scale-Out NIC	ConnectX-8 (800 Gb/s)	ConnectX-9 (1.6 TB/s)
CPU-GPU Interconnect	NVLink-C2C (900 GB/s)	NVLink-C2C (1.8 TB/s)