Comment by LoganDark

IIRC, ease of implementation (for the GPU kernels), and cross-compatibility (the same bytecode can be loaded by multiple models of GPU).