NVIDIA "Vera Rubin" Architecture
The physical limit of the silicon reticle.
Explore this event on the interactive timeline →Unveiled at CES 2026, the Rubin architecture pushes monolithic silicon to its physical limits to satisfy the insatiable demand of AI data centers.
Key Numbers
- Transistors
- 336 Billion (3nm)
- Performance
- 50 petaFLOPS (FP4)
Verified Facts
- Built on a 3nm process, the Rubin GPU features a staggering 336 billion transistors (a 1.6x increase over the 2024 Blackwell B200).
- It delivers approximately 50 petaFLOPS of FP4 inference per GPU, completely redefining what is possible in a single server rack.
- To prevent memory bottlenecks from starving the processor, it utilizes 288 GB of next-generation HBM4 memory with an incredible 22 TB/s bandwidth.