News
Rumored specifications for Rubin GPU architecture include 28,672 CUDA cores and 50 PFLOPS performance
Saturday, January 10, 2026 at 09:00 PM
Preliminary specifications for the upcoming Rubin architecture suggest a configuration of 8 Graphics Processing Clusters (GPCs) and 224 Streaming Multiprocessors (SMs), totaling 28,672 CUDA cores and 896 Tensor cores. The platform is expected to achieve 50 PFLOPS of compute performance through 6th-Gen Tensor cores and increased thermal design power (TDP).
Context
Nvidia is accelerating its data center roadmap with the Rubin architecture, shifting to an aggressive annual release cycle to maintain its lead in AI compute. Rumored specifications for the flagship GPU include 28,672 CUDA cores and 224 Streaming Multiprocessors, representing a 1.4x increase in core density over the current Blackwell series. Powered by 6th-Gen Tensor Cores, the architecture is projected to reach 50 PFLOPS of NVFP4 performance, delivering a 3.3x increase in inference throughput while utilizing advanced TSMC 3nm fabrication.
The platform will also debut HBM4 memory with 22 TB/s of bandwidth and the custom 88-core Vera CPU. By combining these with NVLink 6 interconnects, Nvidia aims to reduce AI token costs tenfold and handle trillion-parameter models more efficiently. Production is expected to ramp in the second half of 2026, targeting the next generation of "AI Factories." This hardware refresh reinforces the company's dominance in the semiconductor supply chain as competitors struggle to match its rapid innovation pace.
Related Companies
Nvidia
NVDA