batch=64: steps=15, modeled_seconds=0.246, examples/s=3902.4
batch=256: steps=15, modeled_seconds=0.534, examples/s=7191.0
batch=1024: steps=15, modeled_seconds=1.686, examples/s=9110.3
More examples/s did not mean less time to target.
noise SD=0.0: best tested batch=16, expected-loss target in 0.1740 modeled s
noise SD=0.2: best tested batch=32, expected-loss target in 0.2508 modeled s
noise SD=0.8: best tested batch=512, expected-loss target in 1.1628 modeled s
