November 22, 11:00 AM, ISCD seminar room (Esclangon building, 1st floor)
Abstract: With emphasis on large-scale dense linear algebra, the static and deterministic dataflow computing paradigm that characterizes Groq's Tensor Streaming Processor (TSP) will be presented. The novelty of this SIMD approach will emerge through an overview of the HW architecture and its scalability. The resulting HPC programming environment will be illustrated with some examples of data analysis and with the simulated adiabatic and chaotic evolution of a classical nonlinear system to solve combinatorial optimization problem in the Qubo/Ising form. Finally, the recent performance with an inference rate of 237 tokens/s on Lama-2 70B, a large language model, will be highlighted.
Curriculum: Ernesto Bonomi graduated as an Engineer in Physics from the Swiss Federal Institute of Technology in Lausanne, Switzerland and completed the Advanced Master's degree in Physics (Doctorat de 3-ème Cycle) at the University of Orléans, France. He was also awarded a Ph.D. in Theoretical Physics (Doctorat ès Science Physique) at the Ecole Polytechnique and University of Paris XI, France.
Former Director at CRS4 of the "HPC for Energy and Environment" Sector. Currently, Adjunct Professor of "Numerical Simulation and HPC" at the Department of Mathematics and Computer Science of the
University of Cagliari, and consultant in the areas of scientific applications of HPC and science education.