Supernode (超节点) technology and AI computing infrastructure research wiki — covering architecture, protocols, standards, and the global competitive landscape of tightly-coupled AI accelerator clusters.
Last updated: 2026-04-20 | 20 sources | 30 wiki pages | HTML Report (V4)
| Metric | Value |
|---|---|
| Domestic supernode market (2028E) | 341.4B RMB, CAGR 194% |
| NVLink 6 bandwidth | 3.6 TB/s per GPU |
| NVL72 3-year TCO | $5.5–6.5M per rack |
| Max chip interconnect | 9,216 (Google Ironwood) |
| Competing Scale-up protocols | 10+ |
| New domestic vendors (2026) | 7+ |
| Vendor | Product | Scale | Interconnect | Status |
|---|---|---|---|---|
| NVIDIA | NVL72 / NVL144 (Vera Rubin) | 72–576 GPU | NVLink 6 (+ Fusion) | NVL72 shipping |
| Ironwood | 9,216 TPU | Optical ICI | Internal production | |
| AMD | MI350 / MI400 Helios | 72 | UALink 2.0 | MI350 shipping, Helios 2026 |
| Huawei | CloudMatrix 384 / Atlas 950 | 384–8,192 NPU | Lingqu / UB-Mesh | CM384 shipping |
| Alibaba | Panjiu AL128 | 128–144 GPU | ALink (multi-protocol) | AL64 deployed |
| China Mobile | OISA | 32–128 | OISA Gen1.1 | Early production |
| Upscale AI | SkyHammer | TBD | UALink switch ASIC | Samples end 2026 |
wiki/
├── index.md # Entry point — read this first
├── overview.md # Research theme and scope
├── log.md # Chronological operation log
├── open-questions.md # 40+ tracking items
├── sources/ # 20 source summaries [1]–[20]
├── concepts/
│ ├── supernode.md # Core concept (20 references)
│ └── scale-up-network.md
├── entities/
│ ├── oaii.md # Open AI Infra Community
│ └── zhixingxing.md # Hard-tech education platform
└── syntheses/
├── 2026-04-20_supernode-report.html # Visual report (V4, 20 citations)
└── 2026-04-20_last30days-supernode-landscape.md
- Protocols: NVLink (Gen 1–6), UALink 2.0, OISA, Google ICI, Huawei HCCS/Lingqu, ALink, OLink, CLink, EthLink, ETH-X, UB-Mesh
- Economics: NVL72 TCO breakdown, distributed vs supernode cost comparison
- Standards: OAII specs, AISBench V3.0, ClusterBench, UALink IP interop testing
- Trends: NVLink Fusion platform strategy, domestic "Year One", 800V DC power, liquid cooling mandatory
- Expert profiles: Jensen Huang, Bill Dally, Lisa Su, Xu Zhijun, Zheng Jiaqi, and others
Securities reports (Huatai, Dongxing), vendor official (NVIDIA, Huawei, Alibaba Cloud, AMD), standards bodies (OAII, ODCC, UALink Consortium), think tanks (CAICT, GCC/Comentropy), tech media (Zhixingxing, The Register, NextPlatform), community discussions.
This repo is built on wiki-kit — a Markdown knowledge base maintained by an LLM. The LLM reads CLAUDE.md for operation rules and follows a structured workflow:
- Ingest: Process raw sources into
wiki/sources/, update concept/entity pages with[[N]]citations, log everything - Query: Consult wiki first, answer with evidence, save reusable analyses to
wiki/syntheses/ - Lint: Detect contradictions, stale content, orphan pages, unpromoted concepts
raw/ # Read-only source materials (human-managed)
wiki/ # Knowledge layer (LLM-managed)
templates/ # Page structure references
skills/ # Multi-role deep research skill (6 roles)