I was considering the usage of lowp varyings for cases where we could potentially afford the lack of precision. Am I right in assuming that there will be a 4x (presumable) savings in bandwidth (write + read) when compared to mediump? Or could this be strongly HW (GPU model/family) dependent?