Graphcore’s IPU could be cheaper and faster than NVIDIA’s A100, though need sharing on-board RAM.

Supports tensorflow, pytorch, paddlepaddle.

pytorch: poptorch, pytorch-lightning(tpu and ipu)

tensorflow:

from tensorflow.python import ipu

Create an IPU distribution strategy

strategy = ipu.ipu_strategy.IPUStrategy()

with strategy.scope():

…

paddlepaddle:

TensorFlow 1 & 2 support with full performant integration with TensorFlow XLA backend

PyTorch support for targeting IPU using the PyTorch ATEN backend

PopART™ (Poplar Advanced Runtime) for training & inference; supports Python/C++ model building plus ONNX model input

Full support for PaddlePaddle

Other frameworks support coming soon

graphcore