NVIDIA's GTC AI conference in San Jose, California, witnessed the unveiling of cutting-edge technologies by NVIDIA CEO Jensen Huang. The event showcased the Blackwell GPU architecture, named after mathematician David H. Blackwell, as the cornerstone of NVIDIA's next-generation AI portfolio.
The Blackwell GPU architecture is designed to meet the demanding computing and bandwidth requirements of current and future AI workloads. With AI model sizes and parameter counts rapidly increasing, the Blackwell GPU offers approximately 20 PetaFLOPS of AI performance, 4X training performance, and 30X inference performance compared to the previous-gen Hopper GPU, all while achieving 25X better energy efficiency.
One of the remarkable achievements of the Blackwell GPU is the combination of two reticle-limited sized dies into a single GPU using the NV-HBI interface fabric, providing 10TB/s of bandwidth between the dies. This configuration allows for 192GB of HBM3e memory with over 8TB/s of peak memory bandwidth and 1.8TB/s of NVLink bandwidth.
The Blackwell GPU, featuring 208 billion transistors and manufactured using TSMC's 4NP process, sets a new standard in AI processing power. NVIDIA also introduced the GB200 Grace Blackwell Superchip, combining a pair of B200 Tensor GPUs with an NVIDIA Grace CPU to power next-gen systems.
Blackwell incorporates advanced technologies such as the 2nd Generation Transformer Engine supporting FP4 precision, 5th Generation NVLink technology doubling performance, and enhancements in Confidential Computing and Decompression Engine for improved security and performance.
NVIDIA also announced the GB200 NVL72 cluster, connecting 36 GB200-powered systems in a single rack for accelerated AI processing. The company introduced new networking products, including the Quantum-X800 InfiniBand and Spectrum-X800 Ethernet platforms, to complement the Blackwell GPUs.
With a wide range of technologies unveiled at GTC, NVIDIA's advancements in AI, robotics, automotive, and networking sectors position the company for continued success. The adoption of Blackwell by major organizations and cloud service providers underscores its significance in driving AI innovation across industries.
Overall, NVIDIA's comprehensive lineup of products and technologies showcased at GTC reaffirms the company's leadership in AI and sets the stage for future breakthroughs in AI-native technologies and networks.
