Would be good to include these benchmarks. While being DL / CV benchmarks they are not the typical convnets, but have some interesting cases that are harder for current tools.

I don't think we should get into benchmarking every tool under the sun, but it'd be good to be fairly comprehensive in terms of the Julia ecosystem, and perhaps have one point of reference from outside. Options:

• Zygote (I'll work on this)
• ReverseDiff (at least for a subset of models where it's convenient)
• PyTorch (represents a good standard for perf outside of Julia)
• DyNet – a little faster than PyTorch but harder to make part of the build
• TensorFlow – unfortunately though it's quite hard to good GPU microbenchmarks for TF though.