NVIDIA Unveils Next-Gen AI Powerhouses: Rubin and Rubin Ultra GPUs with Vera CPUs

NVIDIA's Next-Generation AI Platforms

NVIDIA has unveiled its roadmap for next-generation AI computing platforms at GTC 2025, showcasing the upcoming Rubin and Rubin Ultra GPU architectures alongside new Vera CPUs. These announcements signal a significant leap in AI processing capabilities, scheduled for deployment in 2026 and 2027 1

1

2

3

4

Vera Rubin: The First Step

Set for release in the second half of 2026, the Vera Rubin platform marks NVIDIA's initial foray into this new generation of AI hardware 1

1

4

• GPU: Two reticle-sized chips delivering up to 50 PFLOPs of FP4 performance
• Memory: 288GB of next-gen HBM4 memory
• CPU: 88-core Vera CPU with custom Arm architecture, featuring 176 threads
• Interconnect: Up to 1.8TB/sec NVLINK-C2C

The Vera Rubin NVL144 platform boasts impressive performance metrics:

• 3.6 Exaflops of FP4 inference and 1.2 Exaflops of FP8 training capabilities
• 13TB/sec of HBM4 memory bandwidth with 75TB of fast memory
• NVLINK and CX9 capabilities rated at 260TB/sec and 28.8TB/sec, respectively

Rubin Ultra: Pushing the Boundaries

Following Vera Rubin, NVIDIA plans to launch the even more powerful Rubin Ultra in the second half of 2027 2

2

3

4

• GPU: Four reticle-sized chips offering up to 100 PFLOPs of FP4 performance
• Memory: 1TB of HBM4 memory across 16 HBM sites
• Scale: NVL system expanded from 144 to 576

The Rubin Ultra NVL576 platform promises unprecedented performance:

• 15 Exaflops of FP4 inference and 5 Exaflops of FP8 training capabilities
• 4.6PB/sec of HBM4 memory bandwidth with 365TB of fast memory
• NVLINK and CX9 capabilities increased to 1.5PB/sec and 115.2TB/sec, respectively

Related Stories

Nvidia unveils Vera CPU with 88 Arm cores to power next-generation AI infrastructure

NVIDIA Unveils Roadmap for Next-Gen AI GPUs: Blackwell Ultra and Vera Rubin

Nvidia Unveils Vera Rubin Superchip: Six-Trillion Transistor AI Platform Set for 2026 Production

NVIDIA Unveils Vision for Next-Gen 'AI Factories' with Vera Rubin and Kyber Architectures

Infrastructure and Power Requirements

To support these advanced systems, NVIDIA is developing new infrastructure solutions 2

2

• Kyber: A new rack infrastructure designed for Rubin Ultra
• Power consumption: Up to 600kW per rack for Rubin Ultra systems
• Cooling: Liquid cooling support in Obereon Racks

Impact on AI Computing

These announcements represent a significant advancement in AI computing capabilities:

• Vera Rubin NVL144 offers a 3.3x increase in FP4 inference and FP8 training over the GB300 NVL72 AI server 3

  3

• Rubin Ultra NVL576 provides a 14x increase in the same metrics compared to GB300 NVL72 3

  3

  4

  4

NVIDIA's roadmap demonstrates the company's commitment to pushing the boundaries of AI computing, with each generation offering substantial improvements in performance, memory capacity, and interconnect speeds. These advancements are poised to enable more complex and powerful AI models, potentially revolutionizing various fields that rely on high-performance computing for AI applications.