The problem with GPUs is that they use simd for parallelism, so the shader cores execute multiple threads in lockstep. This is only really fast if there are a huge number of threads that all do essentially the exact same thing, which is the case for shading pixels and doing machine learning stuff and lots of other things, but running minetest this way might be very difficult.
Most game engines are fastest when parallelized to a number of cpu threads, where branches in one thread won't stall all the others