In his GTC 2024 keynote presentation, Nvidia’s CEO announced a suite of enterprise AI products and significant advances in autonomous vehicles and robotics. Among the highlights were the introduction of the new Nvidia DGX SuperPOD, a supercomputer designed to handle the demanding requirements of generative AI at an unprecedented scale.
The NVIDIA DGX SuperPOD is positioned to be a cornerstone in AI supercomputing, harnessing the power of NVIDIA GB200 Grace Blackwell Superchips. These systems feature a powerful combination of NVIDIA CPUs and GPUs designed to facilitate the training and inference of AI models with trillions of parameters. With a liquid-cooled rack-scale architecture, the DGX SuperPOD delivers 11.5 exaflops of AI supercomputing prowess at FP4 precision, paired with a substantial 240 terabytes of high-speed memory that can extend with additional racks for greater scale.
At the heart of each DGX GB200 system are 36 NVIDIA GB200 Superchips, consisting of Grace CPUs and Blackwell GPUs, all interconnected through the fifth-gen NVLink technology. These superchips boast a performance leap up to 30 times faster than their predecessor for certain AI workloads, specifically highlighting large language model inference.
Jensen Huang, the CEO, emphasized the transformative role of NVIDIA’s AI supercomputers in various industries and nations worldwide. The company’s latest offering, the DGX SuperPOD, integrates the most advanced NVIDIA technologies in computation, networking, and software to empower companies with the capacity to develop their AI algorithms.
For users who need vast shared memory spaces for next-gen AI models, the DGX SuperPOD can be configured to connect 576 Blackwell GPUs across multiple DGX GB200 systems through NVLink, magnifying its processing capabilities.
The DGX SuperPOD’s sophisticated architecture includes NVIDIA BlueField-3 DPUs and the Quantum-X800 InfiniBand networking for seamless, ultra-high-speed connections. Furthermore, the fourth-gen NVIDIA SHARP technology enhances In-Network Computing performance by four times, compared to previous iterations.
This supercomputer is designed to minimize downtime with intelligent predictive-management capabilities, monitoring both hardware and software to detect and remedy inefficiencies. This highly autonomous system can also adjust compute resources on-the-fly and perform hardware replacements without unplanned downtime, ensuring that businesses can maintain continuous operation.
In addition to the DGX SuperPOD, NVIDIA revealed the NVIDIA DGX B200 system, tailored for AI model training and inference. The B200 provides an air-cooled, rack-mounted platform for industries with eight Blackwell GPUs and two newest-generation Intel® Xeon® processors, affirming its status as a pursuit for AI Centers of Excellence.
The DGX B200 models showcase superior performance with FP4 precision, yielding up to 15 times faster real-time inference for trillion-parameter models over preceding systems. The inclusion of advanced networking capabilities with NVIDIA ConnectX-7 NICs, BlueField-3 DPUs, and compatibility with NVIDIA’s Quantum-2 InfiniBand and Spectrum-X Ethernet networking platforms ensures high-speed AI operations.
NVIDIA integrates its AI Enterprise software into all DGX platforms, providing robust frameworks for development and deployment. Additionally, NVIDIA offers specialized support for deploying these systems, including assistance with optimization of AI pipelines and infrastructure post-deployment.
The availability of the NVIDIA DGX SuperPOD and DGX B200 systems is slated for release later in the year. Industry professionals and enthusiasts can absorb more details and updates from the GTC event and presentations, aiming to stay abreast of NVIDIA’s path-breaking innovations in AI and robotics.
