To add some context for the graph above. ASIC is the most energy efficient option, since it is specialized hardware designed to run a specific application. On the con side, it is hard to configure, expensive to build and lacks flexibility in terms of algorithmic or requirement changes.

The second most energy efficient option is FPGA, which is a set of configurable gates that allow for some flexibility but still present high implementation challenges (memory allocation, object oriented design).

Then come GPUs which are suitable for applications with high arithmetic intensity and in the bottom of the list is a CPU.

Adding on to this, it makes sense why modern mobile devices all have a diverse set of hardware. Since energy efficiency is as important as versatility for a device you use often but do not charge often, the cost of figuring out when what functionality should go on what hardware and then implementing for that use case is worth the extended battery life (while keeping the device's speed high)