For inference, there is the option to produce ASIC's with the weights hard coded into the chip. They'll be super fast and power efficient, with the huge downs

For inference, there is the option to produce ASIC's with the weights hard coded into the chip. They'll be super fast and power efficient, with the huge downside they aren't reprogrammable.