Through reverse engineering communities (such as the Asahi Linux team and researchers like @dougallj), it is known that the ANE prefers specific tensor dimension multiples (e.g., channel counts must be multiples of 16 or 64) as well as memory layout formats such as NHWC.