On the computational side, there have been confusions about how TPUs and GPUs relate to BERT. BERT base was trained with 4 TPU pods (16 TPU chips) in 4 days and BERT large with 16 TPUs (64 TPU chips) in 4 days. Does this mean only Google can train a BERT model? Does this mean that GPUs are dead? There are two fundamental things to understand here: (1) A TPU is a matrix multiplication engine — it does matrix multiplication and matrix operations, but not much else. It is fast at computing matrix multiplication, but one has to understand that (2) the slowest thing in matrix multiplication is to get the elements from the main memory and load it into the processing unit. In other words, the most expensive part in matrix multiplication is memory loads. Note the computational load for BERT should be about 90% for matrix multiplication. From these facts, we can do a small technical analysis on this topic.
Deep Learning Hardware Limbo
With the release of the Titan V, we now entered deep learning hardware limbo. It is unclear if NVIDIA will be able to keep its spot as the main deep learning hardware vendor in 2018 and both AMD and Intel Nervana will have a shot at overtaking NVIDIA. So for consumers, I cannot recommend buying any hardware right now. The most prudent choice is to wait until the hardware limbo passes. This might take as little as 3 months or as long as 9 months. So why did we enter deep learning hardware limbo just now?