@meta
  v: 1
  route: /the-math
  generated: 2026-06-10T07:02:39.836Z

@intent
  purpose:    Explain the math: weights, KV cache, framework overhead, and concurrency.
  audience:   ai-engineer, ml-researcher, self-hoster
  capability: understand_formula, learn_terminology

@state
  formula: budget = total_vram * (1 - safety) - kv_cache * concurrency - framework_overhead
  thesis: VRAM ~= params * (bits / 8). Everything else is overhead.
  sections[6]: 01 Weight size: the floor, 02 KV cache: the context tax, 03 Framework overhead: the constant, 04 Concurrency: the multiplier, 05 The full equation, 06 What this site does not do
  read_time_min: 14

@actions
  - id: run_calculator
    method: GET
    href: /#calculator
  - id: view_home
    method: GET
    href: /

@context
  > A long-form explainer of how local LLM memory budgeting actually works. The page walks the additions: weight size, KV cache, framework overhead, concurrency. The thesis: most "can my machine run this" tools skip the math; this one shows it. Includes the VRAM tax example on a stock RTX 4090: 24 GB nominal, ~17.4 GB actual weights budget at ctx 8K, conc 1, 15% safety.

@nav
  self:      /the-math
  parents:   [/]
  drilldown: /#calculator