OK, I see. In any case, even if the performance would not match Tensforflow on TPU this is a major achievement. After all, Google made the TPUs specifically for

OK, I see. In any case, even if the performance would not match Tensforflow on TPU this is a major achievement. After all, Google made the TPUs specifically for Tensorflow. I would not have imagined that it was so "easy" to replace LLVM by XLA.