Skip to content

Latest commit

 

History

History
36 lines (34 loc) · 1.67 KB

2024-03-25.md

File metadata and controls

36 lines (34 loc) · 1.67 KB

2024-03-25

Agenda

  • DICOM v4
    • Status? DSC voted to stop discussions (what I heard)
    • Lots of issues raised as part of brainstorming - what is DICOM going to do about it?
      • Should be some communication to the community
    • Whats I heard
      • Some of the issues DICOM just isn't going to address
        • Performance?
      • FUD
        • Fear of incompatibile
        • Uncertainty about whether there would be enough resources
        • Doubt if its possible
      • "Medical imaging is hard, you need $500/hr developers to succeed"
  • nVidia GTC Feedback
  • Sam Altman trying to raise 7T for "faster AI training hardware" (GPUs?)
  • Medical Imaging AI/ML Training with GPUs (Foundation Model / Deep Learning)
    • If you have a large dataset for training and want to use multiple GPUs, how is the data distributed to the different GPUs
    • For large numbers of parameters, you run out of memory so need to spread the training accross connected GPUs
    • Different ways to parallelize
      • Assign different parts of the model to different GPUs
      • Data parallelization - assign subsets of training data to different GPUs, average results from each GPU and update model
      • Pipeline parallelization - different layers of the models are assigned to different GPUs
      • Hybrid of the above three
    • More and more providers are operationalizing AI
      • Duke, Rad Partners, Kaiser