I noticed something interesting in an article about the PowerVR architecture. The PowerVR 6 series has twice as many 16-bit float units than 32-bit float units, while the K1 has only 32-bit units. Now, if you write shaders for PowerVR you will notice that using lowp variables in your shaders makes it go a lot faster. The reason is, of coarse, because lowp (and mediump) use 16-bit floats. But speed isn't the only thing - the 16-bit float units also use less power. So, this power saving ability is simply not available at all on the Tegra K1.