2025-11-28_tpuv7-google-takes-a-swing-at-the::TPU7-0019

① SA Source

  • Source: 開啟完整 SA 文章
  • Section: Datacenter Network (DCN) – Scaling Beyond 9,216 TPUs
  • Line hint: 495

Context Before

Low Latency and Better Locality: The use of direct links between TPUs means that it is possible to achieve much lower latency for TPUs that are physically located close to one another or are reconfigured to connect directly to each other. TPUs that are close to each other also have better data locality.

Datacenter Network (DCN) – Scaling Beyond 9,216 TPUs

Evidence

The Datacenter Network (DCN) is a network separate to ICI that serves the role of both a typical backend and front-end network. It connects across an even larger domain – 147k TPUs in the case of TPUv7 clusters.

Context After

As discussed in our earlier post on Mission Apollo, where Google proposed replacing the Electronic Packet Switch (EPS)-containing spine layer of the traditional “Clos” architecture with Paloma Optical Circuit Switches (OCS) , Google’s DCN consists of an optically switched Datacenter Network Interconnect (DCNI) layer that combines several aggregation blocks, each of which connects several 9,216 TPU ICI clusters.

In 2022, Google’s Apollo project proposed a DCN architecture that described using 136x136 OCS switches for TPUv4 pods with a pod size of 4,096 TPUs. OCS switches at the DCNI layer were organized into 4 Apollo zones, each containing a maximum of 8 racks of 8 OCS switches for a total of 256 OCS switches. When it comes to Ironwood, to support up to 147k TPUv7s on the same network, we hypothesize that the number of ports on the OCS will nearly double as opposed to increasing the maximum number of OCS switches.

② Atomic Claim

TPUv7 DCN 與 ICI 分離,同時扮演 backend/front-end network,可連接約 147k TPUs 的更大 domain。

  • Epistemic Mode: ASSERTED
  • Mapping Status: COMPLETE

③ Semantic Frame

{
  "attribute": "tpuv7_domain_size",
  "context_nodes": [
    {
      "id": "04_knowledge_base/TPUv7",
      "label": "TPUv7"
    },
    {
      "id": "04_knowledge_base/TPU",
      "label": "TPU"
    }
  ],
  "entity": {
    "id": "04_knowledge_base/Datacenter Network",
    "label": "Datacenter Network"
  },
  "frame_type": "ATTRIBUTE",
  "qualifiers": {
    "condition_text": null,
    "numeric_mentions": [
      "147k"
    ],
    "temporal_mentions": []
  },
  "value": {
    "numeric_mentions": [
      "147k"
    ],
    "value_text": "~147k TPUs"
  }
}

④ Canonical Entity Mapping

RoleSurface LabelCanonical Target
entityDatacenter Network04_knowledge_base/Datacenter Network
context_0TPUv7TPUv7
context_1TPUTPU

⑤ Human Review

請在 Properties 逐項確認:

  • 原文 → Atomic Claim 是否忠實
  • Atomic Claim → Semantic Frame 是否忠實
  • Canonical Entity mapping 是否正確
  • Epistemic mode 是否保留原文語氣
  • 最後選擇 review_action

Review state

Markdown 內文不是正式 approval。只有 Apply bridge 寫入的 Decision Ledger event 才是正式決策。