This is NVIDIA’s new GPU - Blackwell NVL72 Rack

NVIDIA’s unveiling of the Blackwell NVL72 rack centers on scale, interconnectivity, and real world use cases that push both gaming and data center workloads. In the early segment, the host highlights the B200 Blackwell super chip assembly, noting that these dies are paired and connected by an ultra-tiny interconnect gap, enabling multiple GPU dies to function as a single unit. The discussion emphasizes that this is not a consumer GPU but a datacenter grade solution designed to deliver vastly higher performance through NVLink interconnects and a tightly integrated rack design. The crew then breaks down the power and cooling requirements, revealing a system that runs on roughly a kilowatt per GPU with liquid cooling, and they explain the architectural intent of the NVL72 rack as a single, massive GPU-like processing entity, assembled from multiple boards and switch units for ultra-low latency and high bandwidth. They also explore the broader ecosystem, including the MGX and HGX design philosophies, and explain how the NVL72 integrates with Nvidia networking hardware for a complete compute, memory, and interconnect stack. As the tour shifts to applications, the host demonstrates potential use cases across healthcare, climate modeling, and game development with NIMS, autonomous data pipelines, and on-device inference, underscoring how AI workloads like protein folding, drug discovery, and in-game animation can scale with specialized hardware. The video wraps with considerations for developers and gamers, noting that Nvidia will support partner systems and reference DGX configurations, while highlighting ongoing demonstrations of real-time speech-to-mouth animation and AI-assisted storytelling through NIMs, ending with a teaser about future demonstrations and community engagement. Across the presentation, the host communicates a sense of awe about the engineering feat, the sheer data throughput, and the potential for accelerating AI training and inference. The narrative ties the hardware capabilities to tangible outcomes: orders of magnitude performance improvements, linear scaling across multi-die GPUs, and the ability to deploy advanced AI services without resorting to bespoke, custom-built infrastructure. The tone remains pragmatic about power requirements and cooling, while optimistic about how this hardware translates into faster drug discovery, smarter medical imaging, and more immersive game development pipelines. Overall, the video positions the NVL72 rack as a landmark in GPU-centric HPC, with layered components from compute blades to high-speed networking that collectively redefine what a single rack can achieve for AI and simulation workloads.

Topics · science_and_technology · data_center · ai · gpu_computing · high_performance_computing · gaming_tech