For example Apple's m3 neural engine is mere 18 TOPS but it’s FP16.
So windows has bigger number but it’s not apple to apple comparison.
Did author test int8 performance?