Prior versions of the image captioning model took three seconds per training step on an Nvidia G20 GPU, but the version open sourced today can do the same task in a quarter of that time, or just 0 ...