Zach Anderson
Sep 18, 2025 15:58
NVIDIA unveils Blackwell, a groundbreaking architecture designed to power AI factories, enhancing AI inference capabilities with unprecedented scale and efficiency.
NVIDIA has introduced its latest innovation, the Blackwell architecture, designed to redefine the landscape of AI inference. This new architecture aims to power AI factories, which are expected to handle the most complex AI models, according to NVIDIA’s blog.
Surging Demand and Model Complexity
The Blackwell architecture is engineered to meet the escalating demand for AI processing power. Today’s AI models, characterized by their vast complexity, often contain hundreds of billions of parameters. Future models are anticipated to exceed a trillion parameters, necessitating robust infrastructure capable of scaling up and out to accommodate these demands.
To address this, Blackwell focuses on scaling up data centers by integrating thousands of computers into cohesive systems, significantly boosting performance and energy efficiency. This approach is pivotal for powering AI factories that serve nearly a billion users weekly.
Today’s Most Challenging Form of Computing
AI inference, recognized as the most challenging computing form today, requires adaptable and scalable infrastructure. NVIDIA’s GB200 NVL72 system exemplifies this, functioning as a single, massive GPU through a symphony of compute, networking, storage, power, and cooling orchestrated by advanced software. This system integrates tens of thousands of Blackwell GPUs, demonstrating the potential of the Blackwell architecture in AI factories.
Birth of a Superchip
The NVIDIA Grace Blackwell superchip, a core component of this architecture, combines two Blackwell GPUs with one NVIDIA Grace CPU. This integration is facilitated by the NVIDIA NVLink technology, which enables seamless communication and memory sharing between the CPU and GPUs, enhancing performance and throughput for AI workloads.
A Backbone That Clears Bottlenecks
The NVLink Switch spine is another critical innovation, designed to eliminate performance bottlenecks by connecting 72 GPUs across 18 compute trays with over 5,000 high-performance copper cables. This infrastructure can move data at a staggering 130 TB/s, exemplifying the architecture’s capacity to handle extreme-scale AI inference.
Building One Giant GPU for Inference
NVIDIA’s GB200 NVL72 system, weighing over one-and-a-half tons and containing more than 600,000 parts, acts as a virtual GPU. This system represents the pinnacle of factory-scale AI inference, where precision and efficiency are paramount.
GB200 NVL72 Everywhere
NVIDIA has deconstructed the GB200 NVL72 system, enabling partners and customers to configure their own NVL72 systems. Manufactured in over 150 factories globally, these systems reflect NVIDIA’s commitment to expanding the reach and capability of AI technologies.
Time to Scale Out
The convergence of tens of thousands of Blackwell NVL72 systems creates AI factories capable of operating as unified entities. NVIDIA’s Spectrum-X Ethernet and Quantum-X800 InfiniBand switches facilitate this integration, ensuring seamless communication and efficiency across data centers.
Opening Lines of Communication
To support AI factories, the NVIDIA BlueField-3 DPUs offload and accelerate non-AI tasks, optimizing networking, storage, and security operations. This enhancement ensures that AI workloads are prioritized, maximizing the efficiency and output of AI factories.
The AI Factory Operating System
NVIDIA Dynamo serves as the operating system for these AI factories, orchestrating and coordinating AI inference requests to optimize productivity and cost-efficiency. It dynamically allocates GPUs across workloads, adapting to user demands and ensuring optimal performance.
In conclusion, NVIDIA’s Blackwell architecture is more than a technological advancement; it’s a transformative platform set to power the future of AI inference, enabling the construction of the world’s largest computing clusters.
Image source: Shutterstock
Credit: Source link
