Graphcore GC200

The GC200 is Graphcore's second-generation Intelligence Processing Unit, announced in July 2020: 59.4 billion transistors on an 823 square millimetre TSMC 7nm die, 1,472 independent IPU-Cores running 8,832 worker threads, and 900 MB of SRAM inside the processor doing the job HBM does on a GPU. There is no DRAM on the chip at all. Graphcore rates it at 250 TFLOPS of FP16.16 and 62.5 TFLOPS of FP32, and the interesting number is not either of those but the on-chip bandwidth, 45 TB/s per processor, which is roughly twenty times what an HBM2e stack of the same era delivered. The architecture bets that a model which fits in SRAM, or which can be split across enough chips to fit, will beat a bigger machine that has to go out to memory. The part ships only inside the 1U IPU-M2000 IPU-Machine, four at a time, and the machines stack into IPU-POD systems of up to 256 of them, so there is no card and no standalone SKU.