Skip to content

TeskesLab/llm-calculator

Repository files navigation

LLM Calc — VRAM & Performance Calculator

Estimate GPU VRAM requirements and generation speed for running large language models locally or in the cloud.

llmcalc.teske.live

Features

  • 174 LLM models with architecture details
  • 135 GPUs (NVIDIA, AMD, Intel, Apple Silicon)
  • Inference & fine-tuning modes
  • Memory breakdown donut chart
  • Speed & throughput estimates
  • Energy cost from user-defined kWh price

Contributing

Adding or Updating GPUs

Edit src/data/gpu-data.json. Each entry follows this format:

{
  "key": "gpu_key_name",
  "label": "GPU Display Name",
  "vendor": "nvidia | amd | apple | intel",
  "type": "discrete | apu",
  "memory": 24,
  "factor": 4.0,
  "tdp_watts": 450,
  "category": "consumer | datacenter | workstation"
}
  • key — unique snake_case identifier used internally
  • label — human-readable name shown in the dropdown
  • vendor — manufacturer: nvidia, amd, apple, or intel
  • typediscrete for dedicated GPUs, apu for integrated / Apple Silicon
  • memory — VRAM in GB (or unified memory for Apple Silicon)
  • factor — relative performance multiplier (RTX 3060 = 1.0 baseline)
  • tdp_watts — typical power draw in watts (used for energy cost calculation)
  • hourly_price_usd — cloud rental price (informational, not used in calculations)
  • categoryconsumer, datacenter, or workstation

The entry will automatically appear in the GPU selector dropdown grouped by vendor.

Adding or Updating Models

Edit src/models-data.json. Field reference:

Field Description
name Display name
slug URL-safe identifier
provider Organization / company
num_of_params Parameter count in billions
architecture dense or moe
hidden_dim_size Hidden layer dimension
num_of_layers Number of transformer layers
attention_structure mha, gqa, mqa, or mla
position_embedding rope, alibi, learned, etc.
num_of_expert_params Expert parameter count (MoE only)
context_length Maximum context window

Modifying Calculation Logic

The VRAM and TPS formulas live in src/engine/calculator.ts. Quantization byte-per-param maps are in src/engine/quantization.ts.

Development

npm install
npm run dev        # Start dev server at localhost:5173
npm run build      # Production build to dist/
npm run build:ci   # CI build (uses dist/ directly)

Tech Stack

React 19, TypeScript, Vite, Mantine UI, recharts

License

MIT — see LICENSE

About

LLM Training / Inference Calculator

Topics

Resources

License

Stars

2 stars

Watchers

0 watching

Forks

Contributors