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

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.