The hardware is good. It's fairly horrendous to program but comparable to other specialised chips and the lowering from XLA / onnx etc worked well enough.
The 20x performance lead over Nvidia's Pascal looked great. Volta beating Pascal by 10x did unrecoverable damage to Graphcore's marketing slides. I think they got the later generations running a bit faster but it was never the order of magnitude over nvidia they started from.
I'm not sure what they're doing these days. The exit to softbank valued the employee shares at $0 so the engineers that had already left got burned. Most of the engineers I knew there have left but there still seems to be a large headcount.
They made a bad bet on the structure of their designs. They were bought by SoftBank last year; they have a new design in the works, but might have trouble delivering it as apparently there's been a bit of a talent exodus. There are several newer companies trying their luck in the area.
they didn't transition to transformers and I focussed mainly on CNN's, which whilst still popular don't have the same financial draw, and weren't easily re-purposable for transformers.
