I think it'll get to a point with quantisation that GPUs that run them will be more FPGA like than graphics renderers. If you quantize far enough things begin to look more like gates than floating point units. At that level a FPGA wouldn't run your model, it would be one your model.