2025-12-04_aws-trainium3-deep-dive-a-potential::TRN3-0022
① SA Source
- Source: 開啟完整 SA 文章
- Section:
Rack Architecture - Line hint:
285
Context Before
Both the Trainium3 NL32x2 Switched and Trainium3 NL72x2 Switched use an all-to-all switched architecture, but the Trainium3 NL72x2 Switched is the rack architecture that is most comparable to Nvidia’s GB200 NVL72 Oberon architecture. Other than both Oberon and Trainium3 NL72x2 Switched using liquid cooling, Trainium3 NL72x2 Switched integrates the CPU into the compute trays just like Nvidia does with Grace and Vera on the same compute tray as the GPUs. By comparison, Trainium NL32x2 Switched still uses a disaggregated CPU node. Like Oberon, Trainium NL72x2 Switched uses cold plates for liquid cooling of the Trainium3 accelerators and the Graviton 4 CPUs. The big difference of Trainium NL72x2 Switched from the Oberon architecture is the use of cross-rack connectivity to increase the scale up world size to span over two racks.
Rack Architecture
Evidence
With each compute tray housing four Trainium3 and one Graviton4 CPU, there are a total of 144 Trainium3s and 36 Graviton 4s across two racks making up the Trainium3 NL72x2 Switched world size.
Context After
② Atomic Claim
NL72x2 每 compute tray 有 4 顆 Trainium3 + 1 顆 Graviton4 CPU；兩 racks 合計 144 Trainium3 + 36 Graviton4。
- Epistemic Mode:
ASSERTED - Mapping Status:
COMPLETE
③ Semantic Frame
{
"frame_type": "NARY_RELATION",
"participants": [
{
"node": {
"id": "04_knowledge_base/Trainium",
"label": "Trainium3"
},
"role": "accelerator"
},
{
"node": {
"id": "04_knowledge_base/Graviton4",
"label": "Graviton4"
},
"role": "cpu"
},
{
"node": {
"id": "04_knowledge_base/Compute Tray",
"label": "Compute Tray"
},
"role": "tray"
}
],
"qualifiers": {
"condition_text": "NL72x2",
"numeric_mentions": [
"4",
"1",
"144",
"36",
"2"
],
"temporal_mentions": []
},
"relation_type": "NL72X2_COMPUTE_COMPOSITION"
}④ Canonical Entity Mapping
| Role | Surface Label | Canonical Target |
|---|---|---|
| accelerator | Trainium3 | Trainium |
| cpu | Graviton4 | Graviton4 |
| tray | Compute Tray | 04_knowledge_base/Compute Tray |
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。