Running AI models without matrix math means far less power consumption—and fewer GPUs?