Repository navigation
AutoLink English Introduction
AIDC AutoLink is an open-source desktop application purpose-built for AI Data Centers (AIDC), covering the full network design lifecycle: Scale-Up fabric topology (NVLink / UALink / UB), Scale-Out network architecture (Fat-Tree / Rail-Optimized / UEC), device selection, PUE energy estimation, oversubscription validation, rack U-space planning, and delivery reports (Excel / PDF). Feed in your server scale and get a production-ready AI data center network design in minutes.
Current Version: V2.9.3 Platforms: Windows / macOS / Linux License: MIT Languages: 简体中文 / English / 日本語 / 한국어 / 繁體中文
Typical scenarios: 1024-GPU training cluster / NVL72 single-rack domain / Huawei CloudMatrix 384 / Ascend super-node / LLM inference cluster — from template to full delivery report in minutes.
Plan GPU-to-GPU fabric (Scale-Up domains) and server-to-server network (Scale-Out) in one workflow:
- Three protocols: NVLink (NVL72), UALink 1.0 (1024-GPU Pod), UB (CloudMatrix 384 single domain)
- Full-mesh fabric: automatic full-mesh edges within each domain, auto domain partitioning (domain_id)
- Rack allocation: one dedicated rack per Scale-Up GPU (amber-colored Scale-Up racks), adjacent rack IDs within a domain
-
Config-driven: JSON
scale_upsection / INI[scale_up]section, backward compatible with legacy fields, graceful degradation for old configs
- Fat-Tree: 2-tier/3-tier adaptive; param, storage, OOB and business networks designed independently
- Rail-Optimized: automatic NVIDIA SuperPOD 8-Rail architecture
- Protocols: IB / RoCE / UEC (Ultra Ethernet, auto-selection of CPO silicon-photonics switches)
- Port naming: full/custom downlink modes; port prefixes auto-resolved from device interface models
- Layer × zone × group anti-overlap auto layout; functional zones for param/storage/business/OOB/Scale-Up
- Box-select multi-node drag, align toolbar (left/right/top/bottom), node deletion with undo
- Edge hover highlight + detail panel (source/target/speed/network type/cable type)
- Layout persisted to
topology.json, preserved on regeneration; per-project data isolation - Dark-mode adaptive
| Rule | Validation |
|---|---|
| V001-V015 | Connectivity, port overflow, device consistency, PUE compliance, etc. |
| V016 | Total server NICs vs Leaf downlink capacity |
| V017 | Optical module package / distance / speed matching |
| V018 | Pod / Scale-Up domain scale sanity |
| V019 | Total facility power vs power supply capacity |
Air-cooled / cold-plate liquid / immersion cooling; IT, cooling and UPS loss breakdown; PUE target check (< 1.25); parametric re-estimation (outdoor temperature, load ratio, UPS efficiency, free cooling).
- Param network (target 1:1), storage (1:1
2:1), business (3:14:1) - PDF/Excel reports read computed values instead of fixed defaults
34 optical modules (100G ~ 1.6T), auto-selected by speed, distance and cable type (DAC / AOC / SR4 / DR4 / FR4 / LR4), with cost estimation.
42U/49U visualization, multi-constraint bin packing (power/U-space/cooling), dedicated rack per Scale-Up GPU, 3-level power color coding (green/yellow/red), 500ms debounced auto-save.
- Excel: connection table / cabling guide (with optical module & cost) / BOM (aggregated by model) / device list / rack table
- PDF report: 9 chapters (overview/architecture/power/optical/cost/validation/device list/oversubscription/racks), no truncation at 2048 servers
- Project name from config; oversubscription from computed values
| Category | Representative Devices |
|---|---|
| GPU Servers | NVIDIA DGX H100/B200/GB300, Huawei Atlas 800T A2, Hygon K100 AI, Cambricon MLU590 |
| Storage Servers | All-flash / hybrid / parallel FS (BeeGFS, IBM Scale, WekaIO, DDN, VAST) |
| Compute Servers | H3C R4900 G7, Huawei 2288H V7, Inspur NF5280M7 |
| Switches | NVIDIA Quantum/Spectrum, Huawei CE16800/CE6881, H3C S9820/S9850/S12500R, Ruijie |
| Optical Modules | 100G ~ 1.6T full speed range |
| Template | Scenario | Scale | Scale-Up |
|---|---|---|---|
| NVL72-单架 | NVIDIA GB200 NVLink domain | 72 GPU | NVLink 72 single domain |
| ualink_1_0_1024 | UALink 1.0 1024-GPU Pod | 1024 GPU | UALink 1024 |
| cloudmatrix_384 | Huawei CloudMatrix 384 | 384 GPU | UB 384 single domain |
| uec_1_0_cluster | UEC 1.0 cluster | 1024 GPU | — |
| SuperPOD-256 | NVIDIA SuperPOD | 256 GPU | — |
| H100-100台 / H100-128台 | NVIDIA H100 training | 100 / 128 GPU | — |
| L20-推理-64 | L20 inference cluster | 64 GPU | — |
| 国产-昇腾-256 | Huawei Ascend 910B | 256 NPU | — |
| 液冷-H100-256 | Liquid-cooled | 256 GPU | — |
| 中型-512 / 大型-1024 / 超大-2048 | Training clusters | 512 / 1024 / 2048 GPU | — |
| 空项目 | Start from scratch | — | — |
Download from Releases:
-
Windows:
AutoLink-Setup-2.9.3-win.exe -
macOS:
AutoLink-2.9.3-mac-x64.dmg/AutoLink-2.9.3-mac-arm64.dmg -
Linux:
AutoLink-2.9.3-linux.AppImage/.deb
- Create project: pick one of 16 templates or start empty (wizard validates each step, inline errors)
- Configure: server scale, network protocol (IB/RoCE/UEC), Scale-Up config, device selection
- Generate topology: one-click backend computation with 19 validation rules
- Review: topology view / PUE / oversubscription / Scale-Up domains / validation results
- Rack planning: auto-placement + manual adjustment, power monitoring
- Deliver: connection table / cabling guide / BOM / device list / rack table / 9-chapter PDF report
- AI training center buildout: plan GPU cluster network (NVIDIA SuperPOD / Huawei Ascend / CloudMatrix)
- Data center expansion: evaluate oversubscription and PUE of existing networks
- Presales support: rapid network design, cabling guide, BOM cost
- Device comparison: compare specs and prices across vendors
- Education: demonstrate Scale-Up / Fat-Tree / Rail-Optimized topology principles
| Version | Highlights |
|---|---|
| V2.0 ~ V2.5 | Core architecture, device library, rack planning, Rail-Optimized, PUE, react-flow rebuild |
| V2.6 ~ V2.7 | Topology persistence, interaction enhancements, dual-stack (Scale-Up/Scale-Out), UEC/UALink templates, updater rebuild |
| V2.8 | Experience & performance: one-shot Excel loading, editable topology, i18n completion |
| V2.9 | Rack planning rebuild, quality engineering, experience polish, capability closing: Scale-Up GA + V016-V019 rules + PDF report fixes + port prefixes + UEC/template consistency |