Inference speed is important because that's what goes to production typically. If you're already waiting hours-days on training, then it probably takes an order of magnitude improvement to make the investment worth it. As a private consumer, my bottleneck to upgrade is my steam library.