NVIDIA ships Rubin NVL8 with an estimated 26% memory-bandwidth cut

Top News · Reported · servethehome.com

If you're sizing a Rubin deployment, the shipping memory and bandwidth figures are the ones to use. NVIDIA's Vera Rubin rollout is underway in Q4 2026, with hyperscalers making some NVL72 systems generally available to their customers.

Here's what changed:

• HGX Rubin NVL8 memory falls from the preliminary 2.3TB to 2.0TB, with an estimated 26% memory-bandwidth reduction versus preliminary specifications.
• Each NVL72 Rubin GPU retains 288GB of HBM4 across eight 36GB stacks.
• NVL72 racks retain 20.7TB of HBM4, with 1,400TB/sec of GPU memory bandwidth.
• NVL72 scale-out networking rises from the preliminary 28.8TB/sec to 32.4TB/sec bidirectional.
• Rubin's listed NVFP4 throughput stays at 50 PFLOPS for inference and 35 PFLOPS for training.

One catch: FP4 throughput is theoretical, not a measured workload result.

Sources: servethehome.com