NVIDIA Rubin & Rubin Ultra Launching Next Year: Next-Gen Vera CPUs, Up to 1 TB HBM4 Memory, 4-Reticle GPUs, 100PF FP4, and 88 CPU Cores

NVIDIA has unveiled its ambitious plans for the upcoming generation of AI architectures with the introduction of the Rubin and Rubin Ultra GPUs, as well as the Vera CPUs, setting the stage for a significant leap in AI computing capabilities.

Introducing NVIDIA’s Rubin & Rubin Ultra GPUs Alongside Vera CPUs – Next-Gen AI Solutions Launching in 2026-2027

This year, NVIDIA has enhanced its Blackwell suite through the release of the Blackwell Ultra platform, which boasts up to 288 GB of HBM3e memory. However, the company is gearing up to push boundaries even further in 2026 with the launch of its innovative CPU and GPU platforms, codenamed Rubin and Vera.

During the GTC event, NVIDIA provided an in-depth look at these groundbreaking platforms expected to debut in late 2026 and into 2027. The first noteworthy offering is the Vera Rubin system, which is designed to scale from NVL72 solutions up to NVL144. This next-gen AI platform will be available in the second half of 2026 and will leverage advanced liquid cooling techniques in its Obereon Racks.

Regarding specifications, the NVIDIA Vera Rubin NVL144 platform will utilize two state-of-the-art chips. The Rubin GPU comprises two Reticle-sized chips, delivering an impressive 50 PFLOPs of FP4 performance and featuring 288 GB of cutting-edge HBM4 memory. This is complemented by an 88-core Vera CPU, which employs a unique Arm architecture, offering 176 threads and an impressive 1.8 TB/s NVLINK-C2C interconnect.

In terms of performance enhancements, the NVIDIA Vera Rubin NVL144 is projected to achieve 3.6 Exaflops of FP4 inference and 1.2 Exaflops for FP8 training. This marks a substantial 3.3x performance increase compared to the GB300 NVL72, with 13 TB/s of HBM4 memory and an impressive 75 TB of fast memory. The architecture also delivers a 60% boost over the GB300, as well as doubling the NVLINK and CX9 capabilities—rated at 260 TB/s and 28.8 TB/s respectively.

The next platform, Rubin Ultra, is scheduled for a 2027 release and will elevate the NVL system from 144 to 576 units. While the CPU architecture remains consistent with its predecessor, the Rubin Ultra GPU will deploy four reticle-sized chips, significantly enhancing performance to a remarkable 100 PFLOPs of FP4 and a total HBM4e capacity of 1 TB distributed across 16 HBM sites.

The NVIDIA Rubin Ultra NVL576 platform promises staggering performance developments, featuring 15 Exaflops for FP4 inference and 5 Exaflops for FP8 training—representing a 14x improvement over the GB300 NVL72. It will deliver 4.6 PB/s of HBM4 memory and offer 365 TB of high-speed memory, realizing an 8x enhancement compared to the GB300. Additionally, it will achieve 12 times the NVLINK and 8 times the CX9 capabilities, rated at up to 1.5 PB/s and 115.2 TB/s respectively.

Source & Images