Bandwidth-aware sizing for LLM serving. Decode speed tracks memory bandwidth, not TFLOPS — the panel sizes VRAM, throughput, and GPU count live.