We describe APEmille, the latest generation of the APE parallel proces
sors. This machine, an evolution of the APE100 concept, is very effici
ent for LGT simulations as well as for a broader class of applications
requiring massive floating point computations. Several new features c
haracterise this evolution. In particular local addressing capabilitie
s are added to all computing nodes. APEmille also exhibits a higher de
gree of integration with a network of workstations acting as a global
host system. An APEmille system in the Teraflops range will be complet
ed in three-four years. The architecture proposed in this paper is bei
ng currently simulated and evaluated.