Dojo appears to be on track for doing something useful in the summer of this year. The question that really matters is when does it become more competitive than a GPU cluster for training. And GPU clusters are getting better so it is a moving target. But that is the goal I've set for the team is the FSD team running our GPU supercluster needs to tell me that they want to use Dojo instead. I don't know when that will be. I can't say success will be 100% there. We want to overestimate our options and underestimate ourselves. But it does seem like we might pass that threshold next year with Dojo.. Dojo is not needed for FSD but it is a cost optimization on training vast amounts of video data. Cost optimization also is a race of improvement. So if you can train models faster and have a short iteration interval then you can make progress faster. Not everything can be distributed to GPUs. There is some element of serialization. And if Dojo is competitive then it seems like the kind of thing where we would offer it to other companies that want to do neural net training. It is very much a neural net training optimization system. In theory, it should be better than a generalized computing platform or say GPUs that were not really intended for pixel shaders, not directly intended for optimizing training of neural networks. Dojo is like giant ASIC optimized for neural net training, especially video. We are not saying for sure that Dojo will succeed. We think it will. We encourage people who think it is an interesting problem, to join Tesla.
