# LLM Roofline (https://llmroofline.dev)

> Deterministic client-side VRAM, KV-cache, Roofline prefill/decode tokens-per-second, and Buy-vs-Cloud-vs-API TCO calculator for open-weight LLMs.

## Core Resources
- Interactive Calculator: https://llmroofline.dev/
- Mathematical Methodology: https://llmroofline.dev/methodology/
- Structured Dataset JSON: https://llmroofline.dev/api/dataset.json
- Open-Source Dataset & DuckDB/CLI Repo: https://github.com/FrancescoPacifico/llm-roofline-db
- Model Directory (32 models): https://llmroofline.dev/models/
- Hardware Directory (28 profiles): https://llmroofline.dev/hardware/

## Key Architectural Coverage
- MoE Active vs Total Memory Bandwidth (Qwen3-235B-A22B, Qwen3-30B-A3B, Llama 4 Maverick/Scout, DeepSeek-R1/V3)
- MLA Compressed KV Cache (DeepSeek-R1/V3) & Hybrid Sliding-Window KV (Gemma 3, Cohere Command A)
- Unified Memory Safe vs Tuned Limits (macOS sysctl iogpu.wired_limit_mb, AMD Strix Halo Linux TTM GTT)
- Prefill Compute Bound (TTFT) vs Decode Memory-Bandwidth Bound Roofline

## Featured Models
- Qwen3 235B-A22B (235.1B, GQA): https://llmroofline.dev/models/qwen3-235b-a22b/
- Qwen3 32B (32.8B, GQA): https://llmroofline.dev/models/qwen3-32b/
- Qwen3 30B-A3B (30.5B, GQA): https://llmroofline.dev/models/qwen3-30b-a3b/
- Qwen3 14B (14.8B, GQA): https://llmroofline.dev/models/qwen3-14b/
- Qwen3 8B (8.2B, GQA): https://llmroofline.dev/models/qwen3-8b/
- Qwen2.5 Coder 32B (32.8B, GQA): https://llmroofline.dev/models/qwen2-5-coder-32b/
- Qwen2.5 72B Instruct (72.7B, GQA): https://llmroofline.dev/models/qwen2-5-72b/
- Qwen2.5 14B Instruct (14.8B, GQA): https://llmroofline.dev/models/qwen2-5-14b/
- Qwen2.5 7B Instruct (7.6B, GQA): https://llmroofline.dev/models/qwen2-5-7b/
- DeepSeek-R1 671B (671B, MLA): https://llmroofline.dev/models/deepseek-r1-671b/
- DeepSeek-V3 671B (671B, MLA): https://llmroofline.dev/models/deepseek-v3-671b/
- DeepSeek-R1 Distill Llama 70B (70.6B, GQA): https://llmroofline.dev/models/deepseek-r1-distill-70b/

## Featured Hardware
- Apple Mac Mini M4 (16 GB) (16 GB, 120 GB/s): https://llmroofline.dev/hardware/mac-mini-m4-16gb/
- Apple Mac Mini M4 (24 GB) (24 GB, 120 GB/s): https://llmroofline.dev/hardware/mac-mini-m4-24gb/
- Apple Mac Mini M4 Pro (48 GB) (48 GB, 273 GB/s): https://llmroofline.dev/hardware/mac-mini-m4-pro-48gb/
- Apple Mac Mini M4 Pro (64 GB) (64 GB, 273 GB/s): https://llmroofline.dev/hardware/mac-mini-m4-pro-64gb/
- Apple MacBook Pro M3 Max (64 GB) (64 GB, 400 GB/s): https://llmroofline.dev/hardware/macbook-pro-m3-max-64gb/
- Apple MacBook Pro M3 Max (128 GB) (128 GB, 400 GB/s): https://llmroofline.dev/hardware/macbook-pro-m3-max-128gb/
- Apple Mac Studio M4 Max (64 GB) (64 GB, 546 GB/s): https://llmroofline.dev/hardware/mac-studio-m4-max-64gb/
- Apple Mac Studio M4 Max (128 GB) (128 GB, 546 GB/s): https://llmroofline.dev/hardware/mac-studio-m4-max-128gb/
- Apple Mac Studio M2 Ultra (192 GB) (192 GB, 800 GB/s): https://llmroofline.dev/hardware/mac-studio-m2-ultra-192gb/
- Apple Mac Studio M3 Ultra (256 GB) (256 GB, 819 GB/s): https://llmroofline.dev/hardware/mac-studio-m3-ultra-256gb/
- Apple Mac Studio M3 Ultra (512 GB) (512 GB, 819 GB/s): https://llmroofline.dev/hardware/mac-studio-m3-ultra-512gb/
- Ryzen AI Max+ 395 Strix Halo (128 GB) (128 GB, 256 GB/s): https://llmroofline.dev/hardware/ryzen-ai-max-395-128gb/