12.6.5 : AI



Tuesday, Mar 17 12:00 AM - 12:40 AM CET : Accelerated HPC+AI Workflow Enables Live-Steering of Vera C. Rubin Observatory and X-ray Free Electron Laser [S81766]
  • Quynh Nguyen, HPC and AI Alliance Manager, NVIDIA
  • Nate Lust, Astronomical software developer, Princeton University


Tuesday, Mar 17 5:00 PM - 5:40 PM CET : Learn how New AI Coding Agents and Tools Unlock GPU Performance for Everyone [S81590]
  • Jackson Marusarz, Technical Product Manager, NVIDIA


Tuesday, Mar 17 6:00 PM - 6:40 PM CET : LLM-Generated CUDA Kernels: Are We There Yet? [S81653]
  • Mark Gabel, Sr. Manager, AI Quality for Development Tools, NVIDIA
  • Mark Saroufim, Software Engineer, Meta Platforms


Wednesday, Mar 18 12:00 AM - 12:40 AM CET : Swallow LLM: Continual Pre-Training and RL for Sovereign AI [S81710]
  • Rio Yokota, Professor, Institute of Science Tokyo, Institute of Integrated Research, Supercomputing Research Center
  • Kazuki Fujii, Graduate Student, Institute of Science Tokyo


Wednesday, Mar 18 5:00 PM - 5:40 PM CET : Best Practices of Multi-Modal and Vision Generation Training in M-Core [S81515]
  • Junyu Wu, Sr. Developer, Tencent Holdings Ltd


Wednesday, Mar 18 10:00 PM - 10:40 PM CET : Accelerate Native Sparse Attention Kernels on Blackwell by CUTLASS Python [S81470]
  • Akash Mehra : Sr. Gen AI Algorithms Engineer, NVIDIA


LIVE : Thursday, Mar 19 6:00 PM - 6:40 PM CET : Advancing Autonomous Vehicles With World Models [S82446]
  • Sanja Fidler, VP, AI Research; Associate Professor, NVIDIA and University of Toronto
  • Autonoous vehicles idea before 2000
  • Directly from sensor data to driver
  • World simulation is critical to train many variance
  • 100 000s hours of recording => with neural gaussian splat no need to render environment anymore => creating many variations of the same senario
  • NVidia Omniverse NuRec
  • Next step generative world models, and edit the scene on the fly with a prompt
  • Cosmos was trained on 20 millions hours of videos
  • Alpa Dream => real time model
  • Takes a frame and the history frames to predict the next one
  • Only two steps of denoising to produce an image (diffusion model)
  • Generate 12 frames at a time
  • Demo of a simulation generated on the fly by alpadream
  • The model is not trained with collision yet
Thursday, Mar 19 12:20 AM - 12:35 AM CET : Building Sovereign AI: Scaling 100B+ Model Training on NVIDIA Blackwell Infrastructure (Presented by Lablup, Inc.) [EX82047]
  • Jeongkyu Shin, CEO, Lablup, Inc.