I think the size of the networks they are training might already be good motivation for developing custom hardware for training. I would expect their training

I think the size of the networks they are training might already be good motivation for developing custom hardware for training. I would expect their training hardware to be something specifically aimed at optimizing memory bandwidth to support distributing training of their “shared” hydra feature. It’s interesting that the shared hydra feature extractor is able to converge as they keep adding more and more output predictions under a training regime of interleaving asynchronous updates to the model from different predictor networks ... Seems to me the formula they are pursuing with custom hardware might be to support a strategy of 1. keep adding more predictions based on same feature 2. Increase the span of time represented by batches used to train the recurrent networks Both pursuits seem very data efficient in terms of the amount of training data they could conceivably collect per unit time of observation ... Custom hardware with a problem specific memory architecture aimed at efficiently supporting training with very large rnn time slices could be developed that’s more about “make it possible to train this proposed model at all” rather than “make it faster/cheaper to train existing common model architectures”. When custom hardware is required to make it possible to train the model they want, the validity of the hardware development cost bet might end up being more about the effectiveness of the model they think they want than it is about maintaining general purpose performance parity vs any off the shelf hardware options ...