SA260324R-0121_r1

SA260324R-0121

① SA Source

• Source: 開啟完整 SA 文章
• Section: GPU and LPU Integration: Attention FFN Disaggregation (AFD)
• Line hint: 137

Context Before

To hide the communication latency of dispatch and combine, we employ ping pong pipeline parallelism.

Evidence

In addition to splitting batches into micro-batches and computation pipelining like standard pipeline parallelism, the tokens dispatched to the LPUs are combined back to the source GPUs, so they ping pong between the GPUs and the LPUs.

Context After

② Atomic Claim

除了像標準 pipeline parallelism 一樣把 batch 切成 micro-batch 並做 computation pipelining，送到 LPU 的 token 還會被 combine 回來源 GPU，因此 token 會在 GPU 與 LPU 之間來回 ping-pong。

• Epistemic Mode: ASSERTED
• Mapping Status: COMPLETE

③ Semantic Frame

{
  "frame_type": "NARY_RELATION",
  "participants": [
    {
      "node": {
        "id": "04_knowledge_base/Pipeline parallelism",
        "label": "pipeline parallelism"
      },
      "role": "parallelism_method"
    },
    {
      "node": {
        "id": "04_knowledge_base/GPU",
        "label": "GPU"
      },
      "role": "compute_endpoint"
    },
    {
      "node": {
        "id": "04_knowledge_base/Groq LPU",
        "label": "LPU"
      },
      "role": "compute_endpoint"
    }
  ],
  "qualifiers": {
    "condition_text": "batch 會切成 micro-batch 並進行 computation pipelining",
    "numeric_mentions": [],
    "temporal_mentions": []
  },
  "relation_type": "PING_PONG_TOKEN_FLOW"
}

④ Canonical Entity Mapping

Role	Surface Label	Canonical Target
parallelism_method	pipeline parallelism	04_knowledge_base/Pipeline parallelism
compute_endpoint	GPU	GPU
compute_endpoint	LPU	04_knowledge_base/Groq LPU

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。