A training flow makes CNNs fit a fixed edge chip’s number rules

A fixed-function edge chip can run a network only if the network follows its built-in number rules. Those rules cover more than bit width: they also determine how values are scaled and when branches can be brought into agreement.

Can a training flow satisfy these linked requirements while keeping image-recognition accuracy close to less constrained methods?

ShiftQAT trains CNNs with the MAX7800x chip’s scaling rules built in, so the resulting networks can meet its fixed numerical contract.

The MAX7800x accelerator cannot run a network unless its scales, arithmetic, and storage obey the chip’s numerical contract. Can training meet those linked constraints while retaining image-recognition accuracy?

ShiftQAT builds the constraints into training. In ImageNet, an image-recognition test whose top-1 score counts correct highest-ranked labels, its 8-bit accuracy drops were comparable to reported drops from less constrained methods.

The tables also show it alone met every listed scaling property, supporting the claim that the tested scaling rules need not bring a large 8-bit accuracy penalty.

The ImageNet accuracy tables use 32-bit biases, so they do not test the hardware’s 8-bit bias requirement. ShiftQAT’s results are single runs, and cited methods can use different reference models and training regimes; the comparison does not isolate the effect of each constraint.