Hosted by Michiel Hochstenbach
Speaker
Gustavo Ramirez Hidalgo, TU/e
Title
Sweeping through arithmetic intensity: using numerical linear algebra to accelerate the simulation of quarks’ interactions and semiconductors
Abstract
Lattice Quantum Chromodynamics (QCD) is the model currently used for the description of the interaction of the sub-nuclear constituents, known as quarks. The Non-Equilibrium Green’s Function (NEGF) method, on the other hand, is a tool that can be used when modelling quantum transport in semiconductors. After a brief introduction to both of these I will, in this talk, proceed to discuss which are their most computationally expensive core components, and how we can deal with them; the concept of arithmetic intensity emerges here as a natural separator of Lattice QCD and NEGF, which also leads to a difference in the type of numerical linear algebra that we can do in both cases. Then, we will discuss some modern techniques used in accelerating the solution of those difficult kernels.
A third method, known as GW, lives in-between the previous two: although typically high in arithmetic intensity, this changes when one starts to view the computational kernels as an f(A)b problem, i.e., as the action on a vector of a function valued on a matrix. The acceleration of such computations is at the center of my current Marie Curie Postdoctoral Fellowship within the department of Mathematics and Computer Science at TU/e.