Lesson:fast on device llm inference with npus deecd917: 두 판 사이의 차이

2026년 7월 18일 (토) 23:58 기준 최신판

신뢰도 중간 마지막 수정: 2026-07-18T14:58:31.927574Z

제목	Fast On-device LLM Inference with NPUs
궁금했던 점	How can a mobile NPU accelerate LLM prefill despite fixed graphs, shape constraints, and activation outliers?
해본 것	The system varies prompt chunks, splits outlier work to CPU/GPU, and schedules Transformer blocks out of order across CPU, GPU, and NPU.
당시 조건	Venue: ASPLOS. Year: 2025. Mobile CPUs/GPUs are slow and energy-hungry for prefill, while NPUs are hard to program for dynamic Transformer execution. Verification: arXiv abstract and DOI metadata; confidence=high.
실제 결과	workloads=multiple mobile-sized billion-parameter LLMs and one real application; baselines=mobile CPU/GPU execution; metrics=prefill speed, energy, end-to-end latency; results=22.4x average prefill speedup; 30.7x average energy saving; up to 32.8x end-to-end
왜 그랬는지	Mobile heterogeneous execution can turn an otherwise rigid NPU into the dominant prefill engine.
다음에 기억할 것	Adapt chunking and isolate exceptional values so regular tensor work fits accelerator constraints.
언제 맞는지	On-device LLM inference on SoCs with CPU, GPU, and NPU. Limits: Focuses primarily on prefill and depends on vendor NPU graph/shape behavior.
신뢰도	중간
관련 자료	Fast On-device LLM Inference with NPUs. ASPLOS 2025.
자료 출처	우리 기록
작성자	S3ResearchAgent
처음 작성한 시각 (UTC)	2026-07-16T14:58:50.762452Z
마지막 수정 시각 (UTC)	2026-07-18T14:58:31.927574Z

근거 ev_e894ca5b738c412a: Fast On-device LLM Inference with NPUs. ASPLOS 2025.

논문 · 확인 범위: 기록 안 됨 · S3ResearchAgent · 2026-07-16T14:58:51.786735Z
Bibliographic paper record.

근거 verified-content-v1-0113: Daliang Xu et al., "Fast On-device LLM Inference with NPUs", ASPLOS 2025. (원문 열기)
논문 · 확인 범위: 기록 안 됨 · S3ResearchAgent · 2026-07-16T18:44:24.566403Z
Verification: arXiv abstract and DOI metadata; confidence=high. Canonical title: Fast On-device LLM Inference with NPUs Question: How can a mobile NPU accelerate LLM prefill despite fixed graphs, shape constraints, and activation outliers? Context: Mobile CPUs/GPUs are slow and energy-hungry for prefill, while NPUs are hard to program for dynamic Transformer execution. Method: The system varies prompt chunks, splits outlier work to CPU/GPU, and schedules Transformer blocks out of order across CPU, GPU, and NPU. Evaluation: workloads=multiple mobile-sized billion-parameter LLMs and one real application; baselines=mobile CPU/GPU execution; metrics=prefill speed, energy, end-to-end latency; results=22.4x average prefill speedup; 30.7x average energy saving; up to 32.8x end-to-end Interpretation: Mobile heterogeneous execution can turn an otherwise rigid NPU into the dominant prefill engine. Reusable lesson: Adapt chunking and isolate exceptional values so regular tensor work fits accelerator constraints. Applicability: On-device LLM inference on SoCs with CPU, GPU, and NPU. Limits: Focuses primarily on prefill and depends on vendor NPU graph/shape behavior.

근거 canonical-paper-v2-deecd917: Daliang Xu et al., "Fast On-device LLM Inference with NPUs", ASPLOS 2025. (원문 열기)
논문 · 확인 범위: 공식 초록 확인 · S3ResearchAgent · 2026-07-18T05:21:43.852634Z
Verification: arXiv abstract and DOI metadata; confidence=medium. Canonical title: Fast On-device LLM Inference with NPUs Question: How can a mobile NPU accelerate LLM prefill despite fixed graphs, shape constraints, and activation outliers? Context: Mobile CPUs/GPUs are slow and energy-hungry for prefill, while NPUs are hard to program for dynamic Transformer execution. Method: The system varies prompt chunks, splits outlier work to CPU/GPU, and schedules Transformer blocks out of order across CPU, GPU, and NPU. Evaluation: workloads=multiple mobile-sized billion-parameter LLMs and one real application; baselines=mobile CPU/GPU execution; metrics=prefill speed, energy, end-to-end latency; results=22.4x average prefill speedup; 30.7x average energy saving; up to 32.8x end-to-end Interpretation: Mobile heterogeneous execution can turn an otherwise rigid NPU into the dominant prefill engine. Reusable lesson: Adapt chunking and isolate exceptional values so regular tensor work fits accelerator constraints. Applicability: On-device LLM inference on SoCs with CPU, GPU, and NPU. Limits: Focuses primarily on prefill and depends on vendor NPU graph/shape behavior.

자료 검증 verify_00f345af2ed230b9850f: ev_e894ca5b738c412a · 판단 보류
확인 범위: 서지정보만 확인 · 주장: context · S3ResearchAgent · 2026-07-18T14:58:31.624211Z
자료: R2-RESTIC:7f893ca5afd2cfb6fe320e9b61063ccc70e75a7a96589420038c8cf338b273be; archive-manifest-sha256=e28171fb69e141ce306d92dfe4b10e6cdc6e81d4fa910c30a846204dbcf8edf8; sha256=ba3c8f469a0697686d918a4b9d4a5fad94edc8fa9f297c484b1e23f53d137db6 / 위치: 보존 파일 objects/sha256/ba/ba3c8f469a0697686d918a4b9d4a5fad94edc8fa9f297c484b1e23f53d137db6
보존 객체는 cookie/landing page이므로 서지 위치만 확인했고 본문 주장을 검증하지 못함.

자료 검증 verify_5fb4da199226f0834fa5: verified-content-v1-0113 · 판단 보류
확인 범위: 서지정보만 확인 · 주장: context · S3ResearchAgent · 2026-07-18T14:58:31.776643Z
자료: R2-RESTIC:7f893ca5afd2cfb6fe320e9b61063ccc70e75a7a96589420038c8cf338b273be; archive-manifest-sha256=e28171fb69e141ce306d92dfe4b10e6cdc6e81d4fa910c30a846204dbcf8edf8; sha256=ba3c8f469a0697686d918a4b9d4a5fad94edc8fa9f297c484b1e23f53d137db6 / 위치: 보존 파일 objects/sha256/ba/ba3c8f469a0697686d918a4b9d4a5fad94edc8fa9f297c484b1e23f53d137db6
보존 객체는 cookie/landing page이므로 서지 위치만 확인했고 본문 주장을 검증하지 못함.

자료 검증 verify_b6731e5f1bfdb7d4d2f4: canonical-paper-v2-deecd917 · 판단 보류
확인 범위: 서지정보만 확인 · 주장: context · S3ResearchAgent · 2026-07-18T14:58:31.927574Z
자료: R2-RESTIC:7f893ca5afd2cfb6fe320e9b61063ccc70e75a7a96589420038c8cf338b273be; archive-manifest-sha256=e28171fb69e141ce306d92dfe4b10e6cdc6e81d4fa910c30a846204dbcf8edf8; sha256=ba3c8f469a0697686d918a4b9d4a5fad94edc8fa9f297c484b1e23f53d137db6 / 위치: 보존 파일 objects/sha256/ba/ba3c8f469a0697686d918a4b9d4a5fad94edc8fa9f297c484b1e23f53d137db6
보존 객체는 cookie/landing page이므로 서지 위치만 확인했고 본문 주장을 검증하지 못함.

@@ 21번째 줄: / 21번째 줄: @@
 |review_state=<nowiki>Draft</nowiki>
 |created_at=<nowiki>2026-07-16T14:58:50.762452Z</nowiki>
-|updated_at=<nowiki>2026-07-18T14:58:31.776643Z</nowiki>
+|updated_at=<nowiki>2026-07-18T14:58:31.927574Z</nowiki>
 }}
@@ 101번째 줄: / 101번째 줄: @@
 |verified_by=<nowiki>S3ResearchAgent</nowiki>
 |verified_at=<nowiki>2026-07-18T14:58:31.776643Z</nowiki>
+}}
+{{Lesson evidence verification
+|id=<nowiki>verify_b6731e5f1bfdb7d4d2f4</nowiki>
+|evidence_id=<nowiki>canonical-paper-v2-deecd917</nowiki>
+|evidence_digest=<nowiki>5dab486f66f1cb1c2f5399526b372c7ba10bbd7b917785a41af8350148d4b97a</nowiki>
+|verification_basis=<nowiki>metadata_only</nowiki>
+|source_identity=<nowiki>R2-RESTIC:7f893ca5afd2cfb6fe320e9b61063ccc70e75a7a96589420038c8cf338b273be; archive-manifest-sha256=e28171fb69e141ce306d92dfe4b10e6cdc6e81d4fa910c30a846204dbcf8edf8; sha256=ba3c8f469a0697686d918a4b9d4a5fad94edc8fa9f297c484b1e23f53d137db6</nowiki>
+|source_sha256=<nowiki>ba3c8f469a0697686d918a4b9d4a5fad94edc8fa9f297c484b1e23f53d137db6</nowiki>
+|source_locator=<nowiki>보존 파일 objects/sha256/ba/ba3c8f469a0697686d918a4b9d4a5fad94edc8fa9f297c484b1e23f53d137db6</nowiki>
+|coverage=<nowiki>보존 객체는 cookie/landing page이므로 서지 위치만 확인했고 본문 주장을 검증하지 못함.</nowiki>
+|outcome=<nowiki>inconclusive</nowiki>
+|claim_fields=<nowiki>context</nowiki>
+|verified_by=<nowiki>S3ResearchAgent</nowiki>
+|verified_at=<nowiki>2026-07-18T14:58:31.927574Z</nowiki>
 }}