Earlier quoted context omitted.
Whether it's substantially better for deep learning depends on whether they chose to neuter fp16 like they did in all other consumer GPUs. The main bottleneck is not actually the size of the model per se (sizes of the models have been mostly going _down_ as of late), it's the size of the batch during training. Both batch size and model size determine the size of the main memory hog: activations that you need to keep…
I wasn't aware that the Jetson TX1 and TX2 supported fp16; that's interesting and good to know. Thanks for the heads up.
IIRC all recent NVIDIA GPUs support fp16, it's just that the perf of fp16 is severely hampered on "consumer" grade hardware, so the memory/perf tradeoff is not viable. I mean, on the one hand I can see why that is: fp16 is not terribly useful in games. But on the other, it's probably nearly the exact same die with a few things disabled here and there to differentiate it from $7K Tesla GPUs, which as an engineer I find super tacky, much like MS deliberately disabling features in Windows Home that don't really need to be disabled.
