Right?

On my Hackintosh (6-core 3.7GHz i7, AMD rx580) I ran `plaidbench --batch-size 16 keras --train mobilenet` on CPU, then on GPU:

CPU: 55.187s (compile), 1321.354s (execution), 1290.38ms (Inference Latency)

GPU: 10.313s (compile), 36.232s (execution), 35.38ms (Inference Latency)

Only about 30 times faster, thanks plaid