High-performance computing is changing rapidly in response to multiple forces. The first is the break-out of intelligence (AI) as a workload of great commercial value which can productively harness enormous computation. The second is the end of silicon scaling, or the beginning of the era of power-limited computing. These two invoke a third; the need for meta-programming environments which can capture learning processes and express massive heterogeneous parallelism. The Intelligence Processor (IPU) is an architectural outline which responds to these forces; a reinvention of accelerated computing from the ground up. This talk will describe and justify the architecture of Graphcore's evolved "Colossus" IPU, and illustrate its value in real applications.