Specifications

Overview

The NVIDIA Rubin GPU is the successor to Blackwell and the compute heart of the Vera Rubin platform, named for the pioneering astronomer Vera Rubin. With 336 billion transistors, 288 GB of HBM4 at 22 TB/s of memory bandwidth, and 50 petaflops of NVFP4 inference compute, Rubin sets a new bar for AI training and inference at gigascale.

Its third-generation Transformer Engine adds hardware-accelerated adaptive compression for efficient low-precision inference, while HBM4 delivers a 2.75× bandwidth improvement over Blackwell's HBM3e at equivalent capacity. Rubin GPUs pair with the NVIDIA Vera CPU — 88 custom Olympus cores with full Armv9.2 compatibility connected over ultra-fast NVLink-C2C — to form the Vera Rubin Superchip.

At the system level, Rubin powers the Vera Rubin NVL72 rack (72 GPUs + 36 Vera CPUs, 3.6 exaFLOPS FP4), the HGX Rubin NVL8 server board for x86 platforms, and DGX SuperPOD reference architectures. NVIDIA confirmed Rubin entered full production in early 2026, with partner systems shipping in the second half of the year.