What specific hardware acceleration is possible? I think TPUs are already as good as you can get for accelerating large MLPs (half of what a LLM is). Is there that much overhead for the attention blocks?