MatLogica HPC library with an option of Automatic Adjoint Differentiation (AAD)

Event type
Seminars
Event date
Venue
online (only)
Speaker
Evgeny Lakshtanov (University of Aveiro)

During this talk I will briefly review principles of parallelization, both vectorization and multithreading. Similarly I will review some basic facts about automatic differentiation. I will then explain a brand new approach Matlogica uses to create HPC kernels. This approach allows  automatic creation of  two kernels, the first one replicates user valuations and the second performs backpropagation algorithm.  No specific knowledge on multithread safety is required from the user. Finally, I will demonstrate some PDE examples and discuss corresponding metrics like performance speed-up and memory consumption.

 
 

 -------------------------------------------------

Zoom Meeting ID: 852 8915 0495

To ask for the Zoom password please contact:

Other