Graphcore, a UK-based semiconductor company that launched its first Intelligence Processing Unit (IPU) for AI acceleration in 2018 has now unveiled its second-generation chip that’s the world’s most complex chip: The Colossus Mk2 GC200 IPU. The Colossus Mk2 surpasses the Nvidia’s A100 which was their most powerful solution towards AI.
Each Colossus Mk2 GC200 IPU consists of 59.4Bn transistors, 250TFlops AI compute and 900MB of in-processor-memory. Each Mk2 contains 8832 separate parallel threads and 1472 independent processor cores.
The Mk2 that is a 7nm IPU powers the IPU-Machine M2000. The M2000 comprises of four Colossus Mk2 GC200 IPU and 5888 independent processor cores. The machine also packs 1 PetaFlop of AI compute and up to 450GB Exchange Memory along with 2.8Tbps IPU-Fabric for ultra-low latency communication all in a slim 1U blade. The machine provides 180TB/s Exchange Memory bandwidth.
The design of IPU-M2000 is so flexible and modular that you can just start with one and scale to thousands. The machine works as a standalone system. The IPU-M2000 can reach the extent of supercomputing when eight are stacked together or racks of 16 tightly interconnected IPU-M2000 are placed in IPU-POD64 systems.
The IPU-POD64 by Graphcore provides building blocks for deploying thousands of machines for large AI/ML problems. According to the Graphcore, the IPU-PODs features ultra-high bandwidth and near-zero latency all thanks to its own IPU-Fabric technology. These IPU-PODs can scale up to 64,000 IPUs and deliver 16 ExaFlops of AI compute.
InvictusTech 