Use a 32-bit floating-point limb size; support 8-bit float