Rumor
Nvidia Rubin architecture memory bandwidth expected to reach 27 TB/s
Thursday, January 15, 2026 at 05:05 PM
Memory bandwidth for the upcoming Rubin architecture is expected to increase from 22 TB/s to approximately 27 TB/s, likely driven by advancements in HBM integration.
Context
Nvidia’s next-generation Rubin architecture is seeing a significant specification boost, with memory bandwidth expectations increasing from the initially announced 22 TB/s to nearly 27 TB/s. This jump follows reports that Nvidia is pressuring memory suppliers for higher HBM4 pin speeds, targeting roughly 13 Gbps. The shift aims to eliminate "memory wall" bottlenecks that currently limit the scaling of trillion-parameter models and complex mixture-of-experts (MoE) architectures.
The Rubin platform, officially unveiled at CES 2026, represents a full-stack integration featuring the Vera CPU and NVLink 6 interconnects. By hitting the 27 TB/s threshold, the architecture will deliver over triple the bandwidth of the current Blackwell series. Fabricated on TSMC’s 3nm process, the Rubin GPU is already in full production, with volume availability from partners expected in the second half of 2026.
This aggressive cadence reinforces Nvidia’s shift to a one-year product cycle, specifically targeting the "Physical AI" and autonomous agentic sectors. The increased throughput is a critical component of the company's strategy to lower inference token costs while maintaining its dominance in data center infrastructure as model complexity continues to surge.
Related Companies
Nvidia
NVDA