You have $300 to build a Nengo accelerator. Post your specs!

In that case, I would look into “one-hot” representation for encoding spikes and I would make every bit to count. I.e. 1 byte allows you to encode neurons coming from 8 neurons. Small things that make the difference and I would try to use reliable frameworks such as caffe/tensorflow.
$300 covers proof of concept therefore don’t attempt to re-invent the wheel.
Bw
Pedro