Telemetry Methodology & Primary Sources
1. AI API Caching & Multi-Turn Context Multiplication
API bills factor uncached vs. cached prefix rates weighted by your Cache Hit %, plus an async Batch API discount toggle (50% on supported frontier endpoints). Multi-turn sessions amplify average input token volume linearly as conversation state replays across turns.
2. Local LLM VRAM & GQA KV-Cache Math
Model weight memory uses empirical GGUF bit-width multipliers (from 16.0 bpw FP16 down to 4.85 bpw Q4_K_M and 2.2 bpw IQ2_XXS). KV cache computes exact Grouped-Query Attention tensors per token: 2 × layers × kv_heads × head_dim × context_length × bytes_per_element. Apple Silicon hosts cap wired GPU pools below physical unified memory (~72%–82%) unless overridden.
3. Storage (TB vs. TiB), ZFS 80% & US EIA kWh Power
Decimal manufacturer drive Terabytes multiply by 0.909495 to yield operating-system Tebibytes (TiB). Synology SHR-1/SHR-2 allocates layer-by-layer parity slices across mismatched drive sizes. Power billing divides DC load by 80+ PSU efficiency and pulls US state residential $/kWh rates and EPA eGRID emission intensities.