Vera Rubin NVL72 vs Blackwell NVL72: Specs Comparison
Vera Rubin NVL72 vs. Blackwell NVL72: Key Specifications
Single-rack comparison — performance, memory, power, and efficiency metrics
Metric
Blackwell NVL72
Vera Rubin NVL72
Fab Process
TSMC 4NP
TSMC 3nm
GPU Memory
(per GPU)
192–270 GB
HBM3e
288 GB
HBM4 +gen
Memory Bandwidth
(per GPU)
8.0 TB/s
22.0 TB/s 2.75×
NVLink Generation
NVLink 5
1.8 TB/s (bi-dir)
NVLink 6
3.6 TB/s (bi-dir) 2×
Inference Performance
(rack total)
720 PFLOPS
(NVFP4)
3,600 PFLOPS
(3.6 Exaflops) 5×
Training Performance
(rack total)
720 PFLOPS
2,520 PFLOPS 3.5×
CPU per Tray
Grace CPU
72-core
Vera CPU
88-core Olympus
Max Rack TDP
120–140 kW
190–230 kW
Tokens per Dollar
(relative efficiency)
1.0× (baseline)
10.0× 90% cost ↓
* Performance figures are NVIDIA's published targets for Vera Rubin (Computex 2026 announcement). Token-per-dollar efficiency is relative to Blackwell NVL72 baseline under comparable inference workloads. Actual results may vary by workload type.