NVIDIA Vera Rubin Driving Performance Per Watt, Lowest Token Cost for Partners Worldwide
Original reporting by NVIDIA Blog
NVIDIA Vera Rubin refers to the company's latest rack-scale AI supercomputer platform, engineered for gigascale, agentic AI workloads. Unveiled with immediate production ramp-up, Vera Rubin is built from the chip to the grid through "extreme co-design," unifying seven distinct chips and five rack trays into a single, optimized system. This integrated approach, centered around the purpose-built NVIDIA Vera CPU, redefines what an AI factory can be, delivering unparalleled performance per watt and drastically lowering the cost per AI token.
Unprecedented Performance
Early benchmarks from partners like CoreWeave are already demonstrating its transformative power. CoreWeave reported Vera Rubin NVL72 achieving 10 times more throughput per megawatt than its predecessor, Grace Blackwell NVL72, a critical metric for power-constrained AI operations. Google Cloud's new A5X instances, powered by Vera Rubin, promise similar efficiency gains and lower inference costs. This leap is supported by advanced networking, including sixth-generation NVLink and Spectrum-X Ethernet, along with innovative cooling solutions that reduce assembly time and water consumption. Major players such as Microsoft, Oracle Cloud Infrastructure, Google Cloud, and CoreWeave are rapidly deploying Vera Rubin, establishing it as the foundation for the next generation of AI infrastructure, including Europe's expanding open-model ecosystem and breakthrough agentic systems.
NVIDIA Vera Rubin’s arrival marks a pivotal moment for AI infrastructure, delivering an unprecedented leap in efficiency and performance through its extreme co-design philosophy. With production now actively ramping and early deployments by industry giants like CoreWeave, Google Cloud, and Microsoft Azure already demonstrating 10x improvements in tokens per megawatt over previous generations, Vera Rubin is immediately redefining the economics of AI factories globally. Its purpose-built architecture, spanning a specialized Vera CPU to advanced liquid-cooled networking, is precisely engineered for the demanding, low-latency requirements of agentic AI workloads and superlearner systems, enabling unprecedented scale and throughput. This meticulous integration, from chip to rack, underscores a new paradigm in hardware optimization critical for the next wave of intelligent agents.
Shaping AI’s Future
This foundational shift in AI compute holds profound implications beyond mere speed. Vera Rubin directly addresses the critical constraints of power consumption and operational cost, enabling wider access to advanced AI capabilities and accelerating the pace of innovation across various sectors. For Europe, it underpins the burgeoning sovereign AI movement, empowering nations to develop and deploy frontier models under local control and regulations, thereby fostering regional autonomy in a strategically vital domain. Its capacity to handle vastly more complex, multi-modal, and autonomous AI agents will unlock new breakthroughs across science, industry, and daily life, from medical discovery to enhanced automation. Vera Rubin is not just a new platform; it is a meticulously crafted ecosystem designed to power the next generation of intelligent systems, ensuring that AI continues its trajectory of rapid evolution while becoming more sustainable, accessible, and strategically impactful worldwide, setting a new benchmark for how AI infrastructure is conceived and delivered.
Frequently asked questions
- What is the NVIDIA Vera Rubin platform and its primary purpose in AI development?
- The NVIDIA Vera Rubin platform is a new rack-scale AI supercomputer engineered for gigascale AI and the agentic era. It features an extreme co-design across multiple chips, including a specialized Vera CPU, to deliver unprecedented performance, efficiency, and lower token cost. Its primary purpose is to accelerate the training and inference of advanced AI models, particularly for power-constrained AI factories and complex agentic workloads.
- How does NVIDIA Vera Rubin improve performance and efficiency compared to prior AI systems?
- NVIDIA Vera Rubin achieves a significant performance leap through extreme co-design across its seven chips, including a specialized Vera CPU and advanced networking. Benchmarks show it delivers up to 10 times more tokens per megawatt than the Grace Blackwell NVL72, significantly reducing token cost. This efficiency is crucial for power-constrained AI factories and demanding agentic workloads, enabling more intelligence within the same power footprint.
- What are the key technical innovations introduced in the NVIDIA Vera Rubin platform?
- Key innovations include the NVIDIA Vera CPU, purpose-built for agentic workloads with 2x single-threaded performance. The platform features a sixth-generation NVLink for over 2x throughput and 3x lower latency, and Spectrum-X Ethernet with co-packaged optics for 5x lower power. Its extreme co-design extends to a cable-free compute tray, reducing assembly time, and an advanced liquid cooling system that enables chiller-free dry-cooler operation, saving millions of gallons of water.