Everything K-culture — comebacks to K-beauty, straight to your inboxGet it in your inbox

METAL MEDIA

juxhinr/bindwidth

128JavaScriptMIT

A calculator that figures out how many servers you actually need to run an LLM in-house, and whether that's cheaper than renting

Bindwidth is an open-source, browser-only calculator that sizes on-premises large language model deployments and estimates total cost of ownership. It identifies which of three technical limits — KV cache memory, combined prefill/decode serving capacity, or session concurrency ceiling — actually constrains a given setup, then prices hardware against that real bottleneck. It runs entirely client-side with no server, database, or telemetry, and compares owned hardware, sovereign rentals, and vendor subscriptions side by side.

What it does

  1. Treats human interactive users and always-on autonomous agents as structurally different workloads with separate peak-load math
  2. Uses six model architecture parameters — total and active parameters, bytes per parameter, KV cache size per token, accepted tokens per step, prefill/decode ratio — to determine GPU count and speed
  3. Calculates three independent constraints (memory, serving throughput, session ceiling) and flags whichever produces the maximum requirement as the true binding limit
  4. Tags every hardware and model data point with a confidence level (measured, published, community, estimated, etc.) so users can see what's verified versus guessed
  5. Built as pure JavaScript functions with no dependencies, backed by a pinned test suite so calculations stay reproducible

Why it matters

It reframes the common 'should we self-host our AI?' debate as a sizing problem rather than a pure budget question, since buying against the wrong constraint wastes capacity while missing the limit that will actually fail under load. This matters for any team weighing private, on-prem AI infrastructure against cloud subscriptions where data governance or cost control is a factor.

Terms in this repo

  • KV cache · temporary memory a model keeps to remember prior conversation context while generating text
  • prefill/decode · the two serving phases: processing the input prompt (prefill) versus generating output tokens one at a time (decode)
  • TCO (total cost of ownership) · the full cost of running hardware, including purchase, electricity, and staff time, not just the sticker price
  • MoE (mixture-of-experts) · a model architecture where only a subset of parameters activate per token, so size and speed are decoupled
  • on-premises · running software on a company's own servers rather than in the cloud

Repository description (English)

Evidence-aware on-prem LLM inference sizing and TCO calculator

Open on GitHub

Trending repos

All repos →

Latest from METAL MEDIA