Nvidia Unveils Rubin Architecture at GTC 2025

Revolutionary AI Hardware Announced at GTC 2025

At the GTC 2025 conference, Nvidia CEO Jensen Huang unveiled the Rubin Architecture, an AI chip, the highly anticipated successor to the Blackwell series. This announcement marks a major breakthrough in AI computing, as Nvidia introduces not only an advanced Rubin GPU but also its first-ever custom ARM-based CPU, Vera.

Set for release in late 2026, with an enhanced Rubin Ultra expected in 2027, these next-generation chips promise unprecedented computational power, featuring HBM4 and HBM4e memory technology for improved efficiency and bandwidth. The unveiling emphasized Nvidia’s increasing focus on “agentic AI”, which enables AI systems to act autonomously, and “physical AI”, designed to power robots and autonomous machines with a deeper understanding of the real world.

The introduction of a custom CPU alongside a powerful GPU signals a strategic shift for Nvidia. Instead of focusing solely on standalone GPUs, the company is now moving towards fully integrated computing solutions optimized specifically for AI workloads. This shift could provide better performance and efficiency, reducing reliance on general-purpose CPUs from third-party manufacturers.

Rubin Architecture: A Leap in AI Processing Power

The Rubin GPU architecture, launching in late 2026, will feature a dual-die design, capable of delivering up to 50 petaFLOPS at FP4 precision and supporting 288GB of HBM4 memory with a 13 TB/s bandwidth. These GPUs will be deployed in Nvidia’s NVL144 rack systems, working alongside the Vera CPUs for optimal AI processing.

The system will integrate Nvidia’s 6th-generation NVLink switch fabric for high-speed interconnectivity, delivering an aggregate bandwidth of 260 TB/s. Additionally, the architecture will support 1.6 Tbps ConnectX-9 Network Interface Cards (NICs), allowing seamless communication and scalability across AI data centers.

A major change in Nvidia’s approach is the use of a two-die GPU design for the base Rubin, compared to Blackwell’s monolithic approach. This design improves manufacturing yield rates, reducing production costs and making high-performance AI chips more accessible.

Introducing Vera: Nvidia’s First Custom ARM-Based CPU

The Vera CPU, Nvidia’s first custom ARM-based processor, represents a significant milestone in AI hardware development. Featuring 88 custom ARM cores and 176 threads per socket, Vera is designed for high-speed AI computation.

With 1.8 TB/s NVLink connectivity, Vera CPUs will seamlessly integrate with Rubin GPUs, enabling faster data transfers and lower latency. Nvidia claims that Vera will deliver twice the performance of the CPUs used in the Grace Blackwell chips, providing more efficient AI processing.

Unlike Nvidia’s previous ARM-based CPUs, which relied on off-the-shelf designs, Vera is built from the ground up to meet the specific needs of AI applications. This level of customization ensures better power efficiency, optimized memory bandwidth, and AI-specific instruction sets, offering a competitive edge over traditional CPUs.

Rubin Ultra: AI Processing on a Whole New Level

Nvidia plans to push AI hardware even further with Rubin Ultra, launching in 2027. This next-level architecture will double the number of GPU dies per package to four, supporting an unprecedented 1TB of HBM4e memory.

Rubin Ultra-powered NVL576 rack systems are projected to deliver:

• 100+ petaFLOPS at FP4 precision
• 15 exaFLOPS of FP4 inference compute
• 5 exaFLOPS of FP8 compute for AI training

These high-performance capabilities cater to large-scale AI models, reducing training times and accelerating inference workloads for advanced AI research, robotics, and scientific computing.

Do you want to boost your career, follow Linkedin. Read latest news here.

Performance Comparison: Blackwell vs. Rubin vs. Rubin Ultra

These improvements highlight Nvidia’s aggressive push to increase computational power, meeting the rapidly growing demands of AI applications across industries.

Redefining AI Computing with Multi-Die GPUs

In a major shift, Nvidia announced that starting with Rubin, multi-die GPUs will be classified as multiple separate GPUs. For example, the two-die Rubin GPU will be counted as two GPUs, while the four-die Rubin Ultra will be considered four GPUs.

This redefinition aligns with the industry’s move toward multi-chiplet designs, which help improve scalability, power efficiency, and manufacturing yields. However, it may also influence how GPU performance is measured and marketed, as Nvidia seeks to redefine industry benchmarks.

Industry Impact and Future Roadmap

Nvidia’s Rubin architecture is expected to revolutionize AI-driven industries, including:

• Autonomous systems & robotics
• Medical research & diagnostics
• Financial modeling & simulations
• Advanced AI-powered applications

Jensen Huang highlighted that AI is at an inflection point, transitioning toward agentic AI (autonomous reasoning systems) and physical AI (AI-powered machines interacting with the real world). These trends require massive computing power, which Rubin aims to provide.

Nvidia has also confirmed an annual release cycle for its AI hardware, ensuring consistent technological advancements. Future architectures, including Feynman (expected in 2028), will continue this trajectory, reinforcing Nvidia’s dominance in AI computing.

Conclusion: The Future of AI is Rubin

With Rubin GPUs and Vera CPUs, Nvidia is taking a bold step toward fully integrated AI computing solutions. By pushing the limits of performance, memory, and efficiency, Nvidia is setting the stage for the next wave of AI-driven breakthroughs.

As AI applications grow in scale and complexity, Nvidia’s Rubin architecture will be at the forefront, enabling researchers, developers, and enterprises to harness AI’s full potential.

🚀 Stay updated on the latest AI advancements with Money Tech News