Comment by junipertea

I found training multiple models on same GPU hit other bottlenecks (mainly memory capacity/bandwidth) fast. I tend to train one model per GPU and just scale the number of computers. Also, if nothing else, we tend to push the models to fit the GPU memory.