LLM VRAM Calculator

Browser-based tool for estimating VRAM requirements and performance when running large language models on various GPU configurations. Pure client-side with no backend dependency.

Features

• VRAM Estimation: Model weights, KV cache, activations, and framework overhead
• 24 Quantization Formats: FP32, FP16, BF16, FP8, MXFP4, INT8, INT4, NF4, GPTQ, AWQ, GGUF Q2-Q8, NVFP4/6
• Measured weights: weight sizes come from published checkpoints per format where one exists, and the results say whether a figure is measured or estimated
• KV Cache Quantization: Independent from weight quantization (FP16, FP8, INT8, INT4)
• Concurrent Users: Size KV cache for 1–65,536 simultaneous sessions; shows per-user tok/s and TTFT, and the maximum sessions that fit at the chosen context
• KV Storage Tier: Park idle sessions on host memory, NVMe or network storage; shows sessions held, resume vs recompute and tier traffic
• Long Context: Sequence lengths up to 1,048,576 tokens (and beyond for models with a larger native context, up to 10,485,760), with a native-context marker and a warning — never a clamp — when the requested length exceeds it
• Multi-GPU Support: Tensor and pipeline parallelism with bandwidth-aware NCCL overhead (NVLink 5/4/3 and NVLink bridges, Infinity Fabric, PCIe 5/4)
• Multi-Node Scaling: Configure GPUs per server and server count, with a selectable scale-out fabric (1.6TbE / 800GbE SONiC RoCEv2, InfiniBand XDR/NDR, 400GbE, 100GbE, or a custom port speed). Tensor parallelism runs inside each server; pipeline parallelism runs across servers
• Configuration rules: impossible combinations are not offered; any correction (including on a shared link) is explained in one notice
• Offloading: CPU/RAM and NVMe offloading simulation; decode reads offloaded bytes over the host link every step (PCIe, NVMe, or Grace NVLink-C2C), and a host-capacity check flags offloaded memory that doesn't fit the host(s)
• Performance Estimation: Tokens/sec from the bytes each decode step reads (active weights plus every session's KV cache) and, on multiple GPUs, all-reduce latency and pipeline fill; prompt-processing time and time-to-first-token modelled from prefill FLOPs (linear weight term plus quadratic attention term), bottleneck analysis (roofline model)
• Configuration Comparison: Save and diff up to 3 configurations side-by-side
• URL Sharing: LZ-String compressed state in URL hash for shareable links
• MoE Architecture: Correct handling of total vs active parameters
• Export: PDF and PowerPoint (PPTX) export of the full analysis
• Dark Mode: Follows OS prefers-color-scheme on first visit, tracks live changes, manual toggle

Supported Hardware

Vendor	GPUs
NVIDIA Datacenter	H100 PCIe/SXM, H200 SXM/NVL, B200, HGX B300 (8-GPU) / GB300 NVL72 (72-GPU rack), A100 PCIe/SXM, L40S, RTX PRO 6000, RTX 6000 Ada
NVIDIA Consumer	RTX 5090, RTX 4090, RTX 3090, DGX Station (GB300 Desktop Superchip), DGX Spark (GB10)
AMD	MI355X, MI350X, MI325X, MI300X
Apple Silicon	M1/M2/M3 Ultra, M1/M2/M3/M4/M5 Max

Plus custom GPU input for any hardware.

Note on the GB300 Desktop Superchip (NVIDIA DGX Station, Dell Pro Max with GB300): the machine exposes 748 GB of coherent memory — 252 GB HBM3e at 7.1 TB/s plus 496 GB LPDDR5X at 396 GB/s. The calculator models a single memory pool, so this entry encodes only the 252 GB HBM3e tier. A model that spills past 252 GB will still load on the real hardware, but it runs at LPDDR5X speed, not the throughput shown here.

Model Database

54 curated models including Gemma 4, Qwen3.8 / Qwen3.6, LLaMA 3.1 / 3.3 / Llama 4, DeepSeek V4 / R1, GLM 5.2 / 4.7, Kimi K3 / K2.x / Linear, Ling 3.0, Mistral Large 3 / Medium 3.5, MiniMax M2.x / M3, Nemotron 3 / 3.5, LFM2.5, GPT OSS, and more. Plus custom model input.

Quick Start

npm install
npm run dev

Open http://localhost:5173 in your browser.

Development

npm run build          # TypeScript check + Vite production build
npm run typecheck      # TypeScript only (tsc -b --force, project-reference aware)
npm run lint           # Biome linting
npm run lint:fix       # Biome auto-fix
npm run format         # Biome format
npm run test           # Vitest watch mode
npm run test:coverage  # Coverage report (75% thresholds on engines/utils)
npm run refresh:all    # Refresh model/GPU data from HuggingFace

Or use make help for all available targets.

Technology Stack

• React 19 + Vite 7
• TypeScript (strict mode, noUncheckedIndexedAccess)
• Zustand (state management with URL hash persistence)
• Tailwind CSS v4 (dark mode)
• Recharts (visualization)
• Decimal.js (precision arithmetic)
• Zod (schema validation, single source of truth for types)
• Web Workers (non-blocking calculations)
• Biome (linting + formatting)
• Vitest (testing)

Project Structure

src/
├── engines/        # Pure calculation logic (VRAM, performance, multi-GPU, offloading)
├── components/
│   ├── inputs/     # Model/GPU selectors, quantization, parameters
│   ├── outputs/    # Charts, tables, fit indicator, recommendations
│   ├── comparison/ # Side-by-side config comparison
│   ├── common/     # Dark mode toggle, shared components
│   └── layout/     # App layout, input/results panels
├── data/           # Static JSON databases (gpus.json, models.json)
├── store/          # Zustand stores (UI state, comparison, URL serialization)
├── hooks/          # useInferenceCalculation, useDarkMode, useURLSync
├── utils/          # Zod schemas, validation helpers
├── types/          # TypeScript type re-exports from Zod
├── workers/        # Web Worker for calculation offloading
└── test/           # Test setup
scripts/            # Data refresh scripts (HuggingFace fetch)

See ARCHITECTURE.md for detailed technical documentation.

License

MIT