About the Session
The arrival of NVIDIA Grace Blackwell platforms marks a paradigm shift in how we approach the most demanding computational challenges. In this session, Dan Ernst will explore the architecture and real-world impact of Grace Blackwell.
The presentation will provide a deep dive into the enhancements of the Blackwell architecture, including the Transformer Engine with native FP4 support, 8 TB/s of HBM memory performance, and the 1.8 TB/s NVLink-C2C interconnect that fuses the Grace CPU and Blackwell GPUs into a single, high-bandwidth logical unit.
Dan will also outline the software ecosystem essential for maximizing this hardware, highlighting CUDA-X libraries, and the NeMo Framework. Attendees will see performance benchmarks demonstrating how the Grace Blackwell platforms deliver speedups on everything from AI training to HPC simulations to a staggering 50x increase in AI factory output compared to previous generations. Finally, the session will showcase notable results in scientific domains enabled by NVIDIA platforms—from climate modeling to complex engineering and drug discovery.