Wow - This seems at first read to be really impressive work. They got scaling laws up to a reasonable size, 2.7B, and also run a few downstream tasks. Would be interesting to see how a comparable model trained by someone else does, to check their scores against those.
They get real (61%!?) memory savings during training, and inference too.
On top of all that, they then go build an FPGA core which is programmed with a custom assembler. And their code is posted and works seamlessly with huggingface transformers?! Absolutely going to test this out.