Fuel Accelerated Computing and Generative AI with NVIDIA Blackwell

Generative AI has elevated computing to a new era, characterized by the staggering capabilities of AI models boasting 10 trillion or more parameters. When AlexNet kicked off the AI boom in 2012, it used 60 million parameters. Just over a decade later, today s complexity has surged over 160,000 fold. These new models can now find cures for cancer, predict extreme weather events, automate robots to perform industrial inspections, and unlock new economic opportunities in every industry. Yet, the journey to harness their full potential presents challenges, notably the vast computational resources and time required for model training. The new extremely large LLMs combined with the need for real-time inference reveals more challenges and complexities of scale, deployment, and operations. NVIDIA Blackwell is the once-in-a-generation platform with the power and energy efficiency needed to effectively train and infer from these models and serve as the foundation for the age of generative AI. Blackwell architecture will be deployed into trillion-dollar markets and democratize real time usage of the new gargantuan models. Training these models needs NVIDIA Blackwell s exaFLOPs of compute. Deploying them requires dozens of Blackwell GPUs to work as a single unified GPU. In this talk we are going to discuss the key NVIDIA Blackwell architectural innovations that supercharge AI training and inference performance and reduce energy use and the cost of ownership.