Attention Mechanism

The physical barrier modern large language models confront when scaling context length is rarely raw floating-point throughput (FLOPs). It is the memory-access wall between high-bandwidth memory (HBM) and on-chip static random-access memory (SRAM). The standard Scaled Dot-Product Attention operator established by Vaswani et al. (2017) incurs an $O(N^2)$ memory-traffic overhead as sequence length $N$ grows, reducing self-attention primarily to a memory-bound workload. Understanding the Transformer requires stripping away anthropomorphic cognitive metaphors and grounding the discussion in the geometry of matrix projections alongside the hardware constraints of modern GPUs.