Comment by Centigonal

wouldn't less compute result in slower inference, rather than worse performance?