24-bit floating points are used internally for calculations on the GPU, however the current code will still emulate that with 32-bit floating points.
In the future we might want to accurately perform the calculations with correct bitness in the future, but for now we just wrap the calculations around this class.