NVIDIA Discloses Rubin GPU Details Featuring 336 Billion Transistors and 10x Agentic AI Performance
NVIDIA has officially disclosed technical details for its next-generation Rubin GPU architecture, which is manufactured on TSMC's 3nm (N3P) process and packs 336 billion transistors. The new architecture delivers a 10x increase in energy-efficiency throughput for agentic AI workloads compared to the previous Blackwell generation. This architectural leap addresses the massive computational and memory bandwidth demands of agentic AI, which requires autonomous decision-making and tool use. By significantly boosting memory bandwidth to 22 TB/s using HBM4, NVIDIA aims to maintain its dominance in the rapidly evolving AI hardware market. The Rubin GPU features two reticle-limit dies joined by NVIDIA's High-Bandwidth Interface (NV-HBI) and includes 8 GPCs, 224 SMs, and 896 Tensor Cores. It also integrates 288 GB of HBM4 memory with a peak bandwidth of 22 TB/s, alongside support for NVLink 6 offering 3600 GB/s of GPU-to-GPU interconnect bandwidth.
## BACKGROUND
Agentic AI refers to advanced AI systems that can autonomously pursue goals, use tools, and execute multi-step workflows rather than just responding to static prompts. In semiconductor manufacturing, the "reticle limit" represents the maximum physical area size that a lithography machine can expose on a single silicon wafer, prompting chipmakers to connect multiple dies using high-speed interfaces like NV-HBI to build larger processors.