Sergey Maydanov - Python math library reinvented. How and why?

Talk description: A GPU accelerated math library brings new challenges in reaching close to native speeds. Traditional pre-compiled kernels approach has certain flaws that does not allow to get close-to-metal performance. Rethinking traditional array library APIs along with bringing just-in-time compilation and linking under-the-hood allows reaching the peak performance. In this talk we will go through a series of examples illustrating how much performance may be left on the table and how these can be addressed with the new approach to the math library design. Bio: Sergey Maydanov has 25+ years of experience in performance math and machine learning libraries targeting both traditional CPUs and GPUs. His area of specialization is computer arithmetic, special functions approximation, and random number generation. In recent years, Sergey is working on nvmath-python, an open-source Python library that brings a pythonic access to NVIDIA native math libraries.