clm_090a0c23c5f146ad8e0b0560adf17197_r1

2025-12-04_aws-trainium3-deep-dive-a-potential::TRN3-0029

① SA Source

• Source: 開啟完整 SA 文章
• Section: NeuronLink PCIe PHY and UALink Switches
• Line hint: 365

Context Before

Source: SemiAnalysis

NeuronLink PCIe PHY and UALink Switches

Evidence

There are three scale up switch generations implemented during the lifecycle of Trainium3, first the 160-lane Scorpio X PCIe 6.0, then the 320-lane Scorpio-X PCIe 6.0, and finally there will be an option to upgrade to a higher radix 72+ port UALink switch.

Context After

The first two switch tray generations adopt a multi-plane scale up switch architecture that does not technically allow full all-to-all communication within the rack without multiple hops. The Gen1 switches will be rather quickly replaced higher bandwidth and higher radix switches. The table below illustrates the scale-up characteristics of the six different combinations across the two rack SKUs and three different switch tray generations.

② Atomic Claim

Trainium3 scale-up switch roadmap 為 160-lane Scorpio X PCIe 6.0 → 320-lane Scorpio-X → 72+ port UALink。

• Epistemic Mode: EXPECTED
• Mapping Status: COMPLETE

③ Semantic Frame

{
  "attribute": "switch_generation_sequence",
  "context_nodes": [
    {
      "id": "04_knowledge_base/PCIe",
      "label": "PCIe"
    },
    {
      "id": "04_knowledge_base/UALink",
      "label": "UALink"
    }
  ],
  "entity": {
    "id": "04_knowledge_base/Trainium",
    "label": "Trainium3"
  },
  "frame_type": "ATTRIBUTE",
  "qualifiers": {
    "condition_text": null,
    "numeric_mentions": [
      "160",
      "320",
      "72+"
    ],
    "temporal_mentions": []
  },
  "value": {
    "numeric_mentions": [
      "160",
      "320",
      "72+"
    ],
    "value_text": "160-lane PCIe 6.0 -> 320-lane PCIe 6.0 -> 72+ port UALink"
  }
}

④ Canonical Entity Mapping

Role	Surface Label	Canonical Target
entity	Trainium3	Trainium
context_0	PCIe	PCIe
context_1	UALink	UALink

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。