2025-12-04_aws-trainium3-deep-dive-a-potential::TRN3-0030

① SA Source

Context Before

Yet another potential parallelism strategy is to use full EP across both racks but plan around the extra hops. This strategy could work well for very sparse models where the model would be unable to implement TP across racks because the d_model dim is too small. So – even the additional latency of the extra hop passing through an immediately connected Trainium3 would be worth it.

Trainium3 NL72x2 Switched on Gen1 Switch Trays

Evidence

There are four planes of 10 160-lane PCIe switches each, 40 total switches on switch trays per rack, and 8 32 Lane PCIe switches on each of the 18 Compute Trays, which amounts to a total of 144 of the smaller PCIe switches on the Compute Trays per rack. This is a grand total of 184 scale-up switches per rack or 368 scale-up switches across the entire scale-up world of 144 Trainium3s.

Context After

trn3 scale up switch roadmap

Source: SemiAnalysis

② Atomic Claim

Trainium3 NL72x2 Gen1 每 rack 合計 184 個 scale-up switches,兩 racks 共 368 switches 對應 144 顆 Trainium3。

  • Epistemic Mode: ASSERTED
  • Mapping Status: PARTIAL

③ Semantic Frame

{
  "attribute": "nl72x2_gen1_switch_count",
  "context_nodes": [
    {
      "id": "04_knowledge_base/PCIe",
      "label": "PCIe"
    }
  ],
  "entity": {
    "id": "04_knowledge_base/Trainium",
    "label": "Trainium3"
  },
  "frame_type": "ATTRIBUTE",
  "qualifiers": {
    "condition_text": "NL72x2 Gen1",
    "numeric_mentions": [
      "4",
      "10",
      "160",
      "40",
      "8",
      "32",
      "18",
      "144",
      "184",
      "368",
      "144"
    ],
    "temporal_mentions": []
  },
  "value": {
    "numeric_mentions": [
      "4",
      "10",
      "160",
      "40",
      "8",
      "32",
      "18",
      "144",
      "184",
      "368",
      "144"
    ],
    "value_text": "184 scale-up switches per rack; 368 across 144-Trainium3 world"
  }
}

④ Canonical Entity Mapping

RoleSurface LabelCanonical Target
entityTrainium3Trainium
context_0PCIePCIe

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。