I think I'd still rather have the hardware area put into tensor cores for the GPU instead of this unit that's only programmable with onnx.