A specific quantization method where both weights (w) and activations (a) are stored as 8-bit integers, providing a good balance between memory savings and model quality.