clm_0cb5953c0a784b5ab774155598e281ca_r3

SA260324R-0104

回到這篇 SA Coverage

① SA Source

• Source: 開啟完整 SA 文章
• Section: GPU and LPU Integration: Attention FFN Disaggregation (AFD)
• Line hint: 93

Context Before

Source: SemiAnalysis

Evidence

Attention and FFN are subsets of operations in a model. In a model forward pass, attention’s output feeds into a token router, and the token router assigns each token to k experts, where each expert is an FFN. Attention and FFN have very different performance properties. During decode phase, the GPU utilization of attention barely improves when scaling batch size due to being bounded by loading KV cache. In contrast, the GPU utilization of FFN scales with batch size comparatively better.

Context After

This is something we have worked with certain hardware vendors and memory companies on with our inference simulator for more than 6 months. ↗

② Atomic Claim

在 decode 階段，attention 的 GPU utilization 在提高 batch size 時幾乎不會改善，因為受到載入 KV cache 的限制。

• Epistemic Mode: ASSERTED
• Mapping Status: PARTIAL

③ Semantic Frame

{
  "attribute": "ATTENTION_UTILIZATION_SCALING",
  "context_nodes": [
    {
      "id": "04_knowledge_base/Decode",
      "label": "decode"
    },
    {
      "id": "04_knowledge_base/Batch size",
      "label": "batch size"
    },
    {
      "id": "04_knowledge_base/KV cache",
      "label": "KV cache"
    }
  ],
  "entity": {
    "id": "04_knowledge_base/GPU",
    "label": "GPU"
  },
  "frame_type": "ATTRIBUTE",
  "qualifiers": {
    "condition_text": "attention 本身在此 claim 沒有 canonical entity",
    "numeric_mentions": [],
    "temporal_mentions": []
  },
  "value": {
    "numeric_mentions": [],
    "value_text": "在 decode 階段，提高 batch size 時 utilization 幾乎不改善，原因是受到 KV cache 載入限制。"
  }
}

④ Canonical Entity Mapping

Role	Surface Label	Canonical Target
entity	GPU	GPU
context_0	decode	Decode
context_1	batch size	04_knowledge_base/Batch size
context_2	KV cache	04_knowledge_base/KV cache

⑤ V2 Draft（需要修正時直接編輯）

Inline repair

只編輯以下三個 Draft 區塊。完成後將 review_action 設為 REQUEST_REPAIR，再手動執行 Repair Preview。
執行腳本前，Draft 不會寫入 Claim Store 或 Graph。

V2 Atomic Claim Draft

在 decode 階段，attention 的 GPU utilization 在提高 batch size 時幾乎不會改善，因為受到載入 KV cache 的限制。

V2 Semantic Frame Draft

{
  "attribute": "ATTENTION_UTILIZATION_SCALING",
  "context_nodes": [
    {
      "id": "04_knowledge_base/Decode",
      "label": "decode"
    },
    {
      "id": "04_knowledge_base/Batch size",
      "label": "batch size"
    },
    {
      "id": "04_knowledge_base/KV cache",
      "label": "KV cache"
    }
  ],
  "entity": {
    "id": "04_knowledge_base/GPU",
    "label": "GPU"
  },
  "frame_type": "ATTRIBUTE",
  "qualifiers": {
    "condition_text": "attention 本身在此 claim 沒有 canonical entity",
    "numeric_mentions": [],
    "temporal_mentions": []
  },
  "value": {
    "numeric_mentions": [],
    "value_text": "在 decode 階段，提高 batch size 時 utilization 幾乎不改善，原因是受到 KV cache 載入限制。"
  }
}

V2 Revision Metadata Draft

{
  "epistemic_mode": "ASSERTED",
  "mapping_status": "PARTIAL"
}

⑥ 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。