clm_41ebea9722fd4db4a9043f87c1d93540_r1

2025-12-04_aws-trainium3-deep-dive-a-potential::TRN3-0016

① SA Source

• Source: 開啟完整 SA 文章
• Section: Trainium3 Server Types and Specifications Overview
• Line hint: 105

Context Before

Source: SemiAnalysis, AWS

Evidence

Despite maintaining 4 stacks of HBM3E, AWS has achieved a 70% increase in memory bandwidth by going from below average pin speeds of 5.7Gbps for Trn2 to 9.6Gbps for Trn3, which is the highest HBM3E pin speeds we’ve seen yet.

Context After

The scale-up bandwidth per Trainium3 chip is doubled vs Trn2 by moving to PCIe Gen 6 which offers 64Gbps per lane (uni-directional) vs the 32Gbps per lane offered by PCIe Gen 5. Trainium3 uses 144 active lanes of PCIe for scale up, which on Gen6 means each Trainium3 supports 1.2 TB/s (uni-directional) of scale-up bandwidth per chip.

Scale-out bandwidth support is doubled to a maximum of 400 Gb/s, but most Trainium3 racks produced will stick with the 200Gb/s per XPU scale-out speed that was used for Trn2.

② Atomic Claim

Trainium3 維持 4 stacks HBM3E，但 pin speed 從 Trn2 的 5.7Gbps 提升至 9.6Gbps，使 memory bandwidth 增加 70%。

• Epistemic Mode: ASSERTED
• Mapping Status: COMPLETE

③ Semantic Frame

{
  "comparison_expression": "4 stacks maintained; 5.7Gbps -> 9.6Gbps; +70% bandwidth",
  "entities": [
    {
      "id": "04_knowledge_base/Trn2",
      "label": "Trn2"
    },
    {
      "id": "04_knowledge_base/Trn3",
      "label": "Trn3"
    },
    {
      "id": "04_knowledge_base/HBM3E",
      "label": "HBM3E"
    }
  ],
  "frame_type": "COMPARISON",
  "metric": "memory_bandwidth_and_pin_speed",
  "operator": "INCREASES",
  "qualifiers": {
    "condition_text": null,
    "numeric_mentions": [
      "4",
      "70%",
      "5.7Gbps",
      "9.6Gbps"
    ],
    "temporal_mentions": []
  }
}

④ Canonical Entity Mapping

Role	Surface Label	Canonical Target
comparison_entity_0	Trn2	Trn2
comparison_entity_1	Trn3	Trn3
comparison_entity_2	HBM3E	HBM3E

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。