Reading turbo3 directly in VEC attention kernel (scalar dequant, no fp16 buffer) saves 5x bandwidth by avoiding fp16 materialization
Scalar math CANNOT beat tensor core MMA even with 5x bandwidth savings from avoiding fp16 materialization. The dequant-to-fp16 + MMA path uses tensor cores (specialized hardware) for the dot product, while native VEC uses scalar FMA instructions. On RTX 3090, tensor core throughput so far exceeds scalar ALU that the 5x bandwidth saving is irrelevant. LESSON: always use tensor cores for dot products when available. The dequant-to-fp16 "round trip" that looks wasteful is actually optimal because it enables hardware-accelerated MMA.