The Letter describes a scalable neural system, HyperNet, based on a pr
obabilistic RAM-based architecture and using a custom VLSI IC. A syste
m using five HyperNet VLSI ICs and capable of realising up to 10 240 n
eurons has been designed, manufactured and demonstrated to have the po
tential to learn more than three orders of magnitude faster than simul
ations on current workstations.