can anyone explain to me the difference between the various GPU types? Namely Tesla K80 vs V100
The best place where to see the difference is to look at the raw specs reported by NVIDIA. In brief, the V100 has more(16GB vs 12GB of K80) & faster memory, more TFLOPS and Tensor Core units. Tensor Core units allow you to exploit the Mixed Precision Training, a recently technique developed by NVIDIA which can make the training 2x faster and with less memory.
Dedicated vs Preemptive for the Standard K80?
Take a look at this from the docs: preemptible vs dedicated
I'm not a huge hardware guy, so what is the difference in terms of speed for training the same network on each variety?
Here's the benchmarks on our instances for some famous ImageNet models: Benchmarking FloydHub instances.
You should have a 4x improvement using the V100, but you have to optimize your model.
Hope it helps.