Switch and networking chips and computer hardware for generative AI to maximize throughput and reduce latency to enable scalability for large multi-GPU systems