2025-11-28_tpuv7-google-takes-a-swing-at-the::TPU7-0013

① SA Source

Context Before

Source: SemiAnalysis, Google

An innovative design of TPU’s liquid cooling is that the flow rate of the coolant is actively controlled by the valves. This enables much more efficient cooling as the flow can be adjusted depending on the amount of workload that each chip has at any given time. Google’s TPU has also long adopted vertical power delivery, in which the VRM modules of the TPUs are on the other side of the PCB board. These VRM modules also require a cold plate for cooling.

Evidence

Overall, the TPU rack design is much simpler than that of the Nvidia Oberon NVL72 design, which has a much higher density and utilizes a backplane to connect GPUs to scale up switches.

Context After

Inter-Chip Interconnect (ICI) – The Key to Expanding Scale-Up World Size

The building block of Google’s ICI scale-up network for TPUv7 is a 4x4x4 3D torus consisting of 64 TPUs. Each 4x4x4 cube of 64 TPUs maps to one physical rack of 64 TPUs. This is an ideal dimension as all 64 TPUs can be connected electrically to each other and still fit in a physical rack.

② Atomic Claim

TPUv7 rack design 比 Nvidia Oberon NVL72 簡單;Oberon density 更高且用 backplane 連 GPU 與 scale-up switches。

  • Epistemic Mode: ASSERTED
  • Mapping Status: COMPLETE

③ Semantic Frame

{
  "comparison_expression": "TPU rack simpler; Oberon denser and backplane-based",
  "entities": [
    {
      "id": "04_knowledge_base/TPU",
      "label": "TPU"
    },
    {
      "id": "04_knowledge_base/NVIDIA Oberon",
      "label": "NVIDIA Oberon"
    },
    {
      "id": "04_knowledge_base/Backplane",
      "label": "Backplane"
    }
  ],
  "frame_type": "COMPARISON",
  "metric": "rack_design_complexity",
  "operator": "TPU_SIMPLER",
  "qualifiers": {
    "condition_text": null,
    "numeric_mentions": [],
    "temporal_mentions": []
  }
}

④ Canonical Entity Mapping

RoleSurface LabelCanonical Target
comparison_entity_0TPUTPU
comparison_entity_1NVIDIA Oberon04_knowledge_base/NVIDIA Oberon
comparison_entity_2BackplaneBackplane

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。