clm_0bca377f98d24e458bfc4aa80bfb1626_r1

2025-12-04_aws-trainium3-deep-dive-a-potential::TRN3-0026

① SA Source

• Source: 開啟完整 SA 文章
• Section: Trainium3 NL72x2 Switched Scale Out Networking
• Line hint: 311

Context Before

Turning to the host CPU, Graviton4 will be the only time to market CPU option for Trainium3 NL72x2 Switched. The CPU can be upgraded later to the next generation of Graviton during the Trainium3’s lifecycle. In theory, x86 CPUs are also supported as these can also interface with other components via PCIe, but we don’t believe that they will plan to have an x86 Trainium3 NL72x2 Switched SKU and will only offer an x86 Trainium NL32x2 Switched SKU. Because Trainium3 uses PCIe 6.0 and Graviton4 uses PCIe 5.0, two PCIe gearboxes must be placed next to the CPU to convert from PCIe 6.0 to PCIE 5.0 for communication between the CPU and GPUs. For CPU memory, 12 DDR5 DIMM slots are placed next to the CPU with DDR5 DIMM modules of 64GB and 128GB capacities to be used for the mainstream SKU. Two 8TB local NVMe drives per compute tray will be used for local storage.

Trainium3 NL72x2 Switched Scale Out Networking

Evidence

Trainium3 NL72x2 Switched will have the same scale-out networking configuration as Trainium3 NL32x2 Switched, namely a choice between 400G or 200G scale-out bandwidth per Trainium3 chip:

Context After

Option 1: Two Nitro-V6 (2*200G) 400Gbps NIC module per JBOG tray of four Trainium3 chips: 200Gbps of EFA bandwidth per Trainium3

② Atomic Claim

Trainium3 NL72x2 與 NL32x2 的 scale-out networking 都提供每 chip 200G 或 400G 選項。

• Epistemic Mode: ASSERTED
• Mapping Status: COMPLETE

③ Semantic Frame

{
  "attribute": "scale_out_bandwidth_options",
  "context_nodes": [
    {
      "id": "04_knowledge_base/Scale-out networking",
      "label": "Scale-out networking"
    }
  ],
  "entity": {
    "id": "04_knowledge_base/Trainium",
    "label": "Trainium3"
  },
  "frame_type": "ATTRIBUTE",
  "qualifiers": {
    "condition_text": "NL32x2/NL72x2",
    "numeric_mentions": [
      "200G",
      "400G"
    ],
    "temporal_mentions": []
  },
  "value": {
    "numeric_mentions": [
      "200G",
      "400G"
    ],
    "value_text": "200G or 400G per Trainium3 chip"
  }
}

④ Canonical Entity Mapping

Role	Surface Label	Canonical Target
entity	Trainium3	Trainium
context_0	Scale-out networking	04_knowledge_base/Scale-out networking

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。