This talks describes recent developments in high performance computing in theoretical physics. Two trends are clearly visible in the attempt to satisfy ever growing computing requirements: one trends concentrates on the efficient use of commercial machines, but at the same time the development of application-driven machines is still a very effective approach in several specific cases.

In this talk, I first review the field at large and then discuss in more details the Janus project, that has developed a high performance parallel machine for Monte Carlo simulations of spin-glass systems. I finally present preliminary results of the Janus collaboration on the dynamics of the Edward-Anderson spin glass and on the critical properties of the Q=4 Potts model.