Not true actually. If by offloading you mean to a device that uses a really small model with 8-bit quantized weights you are not actually solving anything.