Gagik Amirkhanyan

Looks like you are using the default strategy: `tf.distribute.get_strategy()` which places your model (including embeddings) on the CPU (not the TPU). In terms of expected throughput, we see somewhere 1-5...