Your question is Why Transformers Are Memory-Bound. Take a moment with it on the right.
Talk me through your thinking if you like. When you're confident, submit your answer and I'll grade it like a real screen (7/10 or better passes).
Why are transformers memory-bound?
Explain the distinction between compute-bound and memory-bound workloads in transformer training and inference. Cover the sources of memory traffic, including parameter reads, activation reads and writes, attention intermediates, optimizer state, and key-value caches. Discuss how sequence length, batch size, precision, fusion, quantization, and distributed execution change the bottleneck. Propose practical methods to diagnose and mitigate memory bandwidth and capacity limits.