NVIDIA HGX Rubin NVL8

8x Rubin Rubin baseboard 2026

Overview

140

FP8 PFLOPS Dense

2.3 TB

Aggregate Memory

176.0 TB/s

Aggregate Bandwidth

—

Aggregate Power

Performance Metrics

Precision	Peak (PFLOPS)	Per GPU (TFLOPS)	TFLOPS/W
FP8	140.0	17500.0	—
NVFP4	280.0	35000.0	—
FP6	140.0	17500.0	—

All figures are dense. Vendors commonly headline the sparse number, which is twice the dense one.

Configuration

Accelerator: Rubin SXM
Count: 8
Memory per GPU: 288 GB
Bandwidth per GPU: 22.0 TB/s

Power & Efficiency

Aggregate Power: —
TDP per GPU: —
FP8 Efficiency: —
Aggregate Bandwidth: 176.0 TB/s

System Details

Vendor: NVIDIA

Family: HGX

Architecture: Rubin

Form Factor: baseboard

Release Year: 2026

Process Node: —

Interconnect

NVLink 6 Switch, 28.8 TB/s total

Notes

Eight Rubin SXM on an HGX baseboard, listed by NVIDIA alongside HGX B300 and HGX B200. NVIDIA publishes NVFP4 twice and the two are different workloads, not a sparse/dense pair: inference 400 PFLOPS sparse, training 280 PFLOPS dense. FP8/FP6 training is 140 PFLOPS dense. System power is not published for the baseboard; the 24 kW figure on DGX Rubin NVL8 is the full chassis including CPUs. NVIDIA marks the specification preliminary.