> In our experiments, we found that we are able to perform int8 quantization on the pseudo gradients without any impact on the loss curves.

Allegedly not?