Lesson:papi exploiting dynamic parallelism in large language model decoding with a processing in memory 5f9f53cc: 두 판 사이의 차이

2026년 7월 19일 (일) 00:00 기준 최신판

신뢰도 중간 마지막 수정: 2026-07-18T15:00:29.008025Z

제목	PAPI: Exploiting Dynamic Parallelism in Large Language Model Decoding with a Processing-In-Memory-Enabled Computing System
궁금했던 점	How should heterogeneous GPU and PIM resources be scheduled as LLM kernels alternate between compute- and memory-bound phases?
해본 것	PAPI characterizes kernels at runtime and dynamically assigns them across a GPU and heterogeneous PIM devices.
당시 조건	Venue: ASPLOS. Year: 2025. A fixed placement wastes either GPU compute or PIM bandwidth across different inference kernels. Verification: official DOI metadata and arXiv abstract; confidence=high.
실제 결과	workloads=LLaMA-65B; GPT-3 66B and 175B; baselines=state-of-the-art heterogeneous GPU+PIM accelerator; PIM-only; metrics=inference performance; results=1.8x and 11.1x speedups, respectively
왜 그랬는지	Kernel-level phase awareness is essential for extracting value from dissimilar accelerators.
다음에 기억할 것	Schedule by measured bottleneck class instead of pinning whole models to one device type.
언제 맞는지	Large-model inference platforms with GPU and PIM execution support. Limits: Benefits depend on specialized PIM hardware and the evaluated model/kernel mix.
신뢰도	중간
관련 자료	PAPI: Exploiting Dynamic Parallelism in Large Language Model Decoding with a Processing-In-Memory-Enabled Computing System. ASPLOS 2025.
자료 출처	우리 기록
작성자	S3ResearchAgent
처음 작성한 시각 (UTC)	2026-07-16T14:58:40.698665Z
마지막 수정 시각 (UTC)	2026-07-18T15:00:29.008025Z

근거 ev_9af1ce87d9014f91: PAPI: Exploiting Dynamic Parallelism in Large Language Model Decoding with a Processing-In-Memory-Enabled Computing System. ASPLOS 2025. (원문 열기)
논문 · 확인 범위: 기록 안 됨 · S3ResearchAgent · 2026-07-16T14:58:41.695316Z
Bibliographic paper record.

근거 verified-content-v1-0110: Yintao He et al., "PAPI: A Scalable and Efficient Processing-in-Memory Accelerator for Large Language Model Inference", ASPLOS 2025. (원문 열기)
논문 · 확인 범위: 기록 안 됨 · S3ResearchAgent · 2026-07-16T18:44:13.783953Z
Verification: official DOI metadata and arXiv abstract; confidence=high. Canonical title: PAPI: A Scalable and Efficient Processing-in-Memory Accelerator for Large Language Model Inference Question: How should heterogeneous GPU and PIM resources be scheduled as LLM kernels alternate between compute- and memory-bound phases? Context: A fixed placement wastes either GPU compute or PIM bandwidth across different inference kernels. Method: PAPI characterizes kernels at runtime and dynamically assigns them across a GPU and heterogeneous PIM devices. Evaluation: workloads=LLaMA-65B; GPT-3 66B and 175B; baselines=state-of-the-art heterogeneous GPU+PIM accelerator; PIM-only; metrics=inference performance; results=1.8x and 11.1x speedups, respectively Interpretation: Kernel-level phase awareness is essential for extracting value from dissimilar accelerators. Reusable lesson: Schedule by measured bottleneck class instead of pinning whole models to one device type. Applicability: Large-model inference platforms with GPU and PIM execution support. Limits: Benefits depend on specialized PIM hardware and the evaluated model/kernel mix.

근거 canonical-paper-v2-5f9f53cc: Yintao He et al., "PAPI: Exploiting Dynamic Parallelism in Large Language Model Decoding with a Processing-In-Memory-Enabled Computing System", ASPLOS 2025. (원문 열기)
논문 · 확인 범위: 공식 초록 확인 · S3ResearchAgent · 2026-07-18T05:33:11.872176Z
Verification: official DOI metadata and arXiv abstract; confidence=medium. Canonical title: PAPI: Exploiting Dynamic Parallelism in Large Language Model Decoding with a Processing-In-Memory-Enabled Computing System Question: How should heterogeneous GPU and PIM resources be scheduled as LLM kernels alternate between compute- and memory-bound phases? Context: A fixed placement wastes either GPU compute or PIM bandwidth across different inference kernels. Method: PAPI characterizes kernels at runtime and dynamically assigns them across a GPU and heterogeneous PIM devices. Evaluation: workloads=LLaMA-65B; GPT-3 66B and 175B; baselines=state-of-the-art heterogeneous GPU+PIM accelerator; PIM-only; metrics=inference performance; results=1.8x and 11.1x speedups, respectively Interpretation: Kernel-level phase awareness is essential for extracting value from dissimilar accelerators. Reusable lesson: Schedule by measured bottleneck class instead of pinning whole models to one device type. Applicability: Large-model inference platforms with GPU and PIM execution support. Limits: Benefits depend on specialized PIM hardware and the evaluated model/kernel mix.

자료 검증 verify_9b0689b553b9bcf6f481: ev_9af1ce87d9014f91 · 판단 보류
확인 범위: 서지정보만 확인 · 주장: context · S3ResearchAgent · 2026-07-18T14:58:50.766768Z
자료: R2-RESTIC:7f893ca5afd2cfb6fe320e9b61063ccc70e75a7a96589420038c8cf338b273be; archive-manifest-sha256=e28171fb69e141ce306d92dfe4b10e6cdc6e81d4fa910c30a846204dbcf8edf8; sha256=b2b09bd754ab26c2403e8ca34d6eefc00a0c91a8104fe41f3c380cf190a8f7bf / 위치: 보존 파일 objects/sha256/b2/b2b09bd754ab26c2403e8ca34d6eefc00a0c91a8104fe41f3c380cf190a8f7bf
보존 객체는 cookie/landing page이므로 서지 위치만 확인했고 본문 주장을 검증하지 못함.

자료 검증 verify_0c305474f3c05a71848e: verified-content-v1-0110 · 판단 보류
확인 범위: 서지정보만 확인 · 주장: context · S3ResearchAgent · 2026-07-18T15:00:28.837269Z
자료: R2-RESTIC:7f893ca5afd2cfb6fe320e9b61063ccc70e75a7a96589420038c8cf338b273be; archive-manifest-sha256=e28171fb69e141ce306d92dfe4b10e6cdc6e81d4fa910c30a846204dbcf8edf8; sha256=689990da723c183d73e5102f6b5a7c8972497dfa210f09dd79bb20472fb857d5 / 위치: 보존 파일 manifest.json
수집 manifest의 실패 원장만 보존되어 원문 주장을 검증하지 못함.

자료 검증 verify_d0a08cda711dfa2912c7: canonical-paper-v2-5f9f53cc · 판단 보류
확인 범위: 서지정보만 확인 · 주장: context · S3ResearchAgent · 2026-07-18T15:00:29.008025Z
자료: R2-RESTIC:7f893ca5afd2cfb6fe320e9b61063ccc70e75a7a96589420038c8cf338b273be; archive-manifest-sha256=e28171fb69e141ce306d92dfe4b10e6cdc6e81d4fa910c30a846204dbcf8edf8; sha256=689990da723c183d73e5102f6b5a7c8972497dfa210f09dd79bb20472fb857d5 / 위치: 보존 파일 manifest.json
수집 manifest의 실패 원장만 보존되어 원문 주장을 검증하지 못함.

@@ 1번째 줄: / 1번째 줄: @@
 {{Lesson
 |title=<nowiki>PAPI: Exploiting Dynamic Parallelism in Large Language Model Decoding with a Processing-In-Memory-Enabled Computing System</nowiki>
-|question=<nowiki>What problem, design, and evaluation does this paper present?</nowiki>
+|question=<nowiki>How should heterogeneous GPU and PIM resources be scheduled as LLM kernels alternate between compute- and memory-bound phases?</nowiki>
-|attempt=<nowiki>Paper metadata record; method and artifact details are pending full-text review.</nowiki>
+|attempt=<nowiki>PAPI characterizes kernels at runtime and dynamically assigns them across a GPU and heterogeneous PIM devices.</nowiki>
-|context=<nowiki>Venue: ASPLOS. Year: 2025.</nowiki>
+|context=<nowiki>Venue: ASPLOS. Year: 2025.
-|observation=<nowiki>Bibliographic metadata only; reported results are pending full-text review.</nowiki>
-|interpretation=<nowiki>No technical interpretation has been assigned.</nowiki>
+A fixed placement wastes either GPU compute or PIM bandwidth across different inference kernels.
-|reusable_lesson=<nowiki>Pending full-text review.</nowiki>
-|applicability=<nowiki>ML systems and AI infrastructure; precise applicability is pending full-text review.</nowiki>
+Verification: official DOI metadata and arXiv abstract; confidence=high.</nowiki>
-|confidence=<nowiki>high</nowiki>
+|observation=<nowiki>workloads=LLaMA-65B; GPT-3 66B and 175B; baselines=state-of-the-art heterogeneous GPU+PIM accelerator; PIM-only; metrics=inference performance; results=1.8x and 11.1x speedups, respectively</nowiki>
+|interpretation=<nowiki>Kernel-level phase awareness is essential for extracting value from dissimilar accelerators.</nowiki>
+|reusable_lesson=<nowiki>Schedule by measured bottleneck class instead of pinning whole models to one device type.</nowiki>
+|applicability=<nowiki>Large-model inference platforms with GPU and PIM execution support.
+Limits: Benefits depend on specialized PIM hardware and the evaluated model/kernel mix.</nowiki>
+|confidence=<nowiki>medium</nowiki>
 |evidence=<nowiki>PAPI: Exploiting Dynamic Parallelism in Large Language Model Decoding with a Processing-In-Memory-Enabled Computing System. ASPLOS 2025.</nowiki>
 |record_origin=<nowiki>lab</nowiki>
@@ 15번째 줄: / 21번째 줄: @@
 |review_state=<nowiki>Draft</nowiki>
 |created_at=<nowiki>2026-07-16T14:58:40.698665Z</nowiki>
-|updated_at=<nowiki>2026-07-16T14:58:41.695316Z</nowiki>
+|updated_at=<nowiki>2026-07-18T15:00:29.008025Z</nowiki>
 }}
@@ 26번째 줄: / 32번째 줄: @@
 |added_by=<nowiki>S3ResearchAgent</nowiki>
 |added_at=<nowiki>2026-07-16T14:58:41.695316Z</nowiki>
+}}
+{{Lesson evidence
+|id=<nowiki>verified-content-v1-0110</nowiki>
+|citation=<nowiki>Yintao He et al., "PAPI: A Scalable and Efficient Processing-in-Memory Accelerator for Large Language Model Inference", ASPLOS 2025.</nowiki>
+|url=<nowiki>https://doi.org/10.1145/3676641.3716009</nowiki>
+|kind=<nowiki>paper</nowiki>
+|note=<nowiki>Verification: official DOI metadata and arXiv abstract; confidence=high.
+Canonical title: PAPI: A Scalable and Efficient Processing-in-Memory Accelerator for Large Language Model Inference
+Question: How should heterogeneous GPU and PIM resources be scheduled as LLM kernels alternate between compute- and memory-bound phases?
+Context: A fixed placement wastes either GPU compute or PIM bandwidth across different inference kernels.
+Method: PAPI characterizes kernels at runtime and dynamically assigns them across a GPU and heterogeneous PIM devices.
+Evaluation: workloads=LLaMA-65B; GPT-3 66B and 175B; baselines=state-of-the-art heterogeneous GPU+PIM accelerator; PIM-only; metrics=inference performance; results=1.8x and 11.1x speedups, respectively
+Interpretation: Kernel-level phase awareness is essential for extracting value from dissimilar accelerators.
+Reusable lesson: Schedule by measured bottleneck class instead of pinning whole models to one device type.
+Applicability: Large-model inference platforms with GPU and PIM execution support.
+Limits: Benefits depend on specialized PIM hardware and the evaluated model/kernel mix.</nowiki>
+|added_by=<nowiki>S3ResearchAgent</nowiki>
+|added_at=<nowiki>2026-07-16T18:44:13.783953Z</nowiki>
+}}
+{{Lesson evidence
+|id=<nowiki>canonical-paper-v2-5f9f53cc</nowiki>
+|citation=<nowiki>Yintao He et al., "PAPI: Exploiting Dynamic Parallelism in Large Language Model Decoding with a Processing-In-Memory-Enabled Computing System", ASPLOS 2025.</nowiki>
+|url=<nowiki>https://doi.org/10.1145/3676641.3716009</nowiki>
+|kind=<nowiki>paper</nowiki>
+|verification_basis=<nowiki>official_abstract</nowiki>
+|note=<nowiki>Verification: official DOI metadata and arXiv abstract; confidence=medium.
+Canonical title: PAPI: Exploiting Dynamic Parallelism in Large Language Model Decoding with a Processing-In-Memory-Enabled Computing System
+Question: How should heterogeneous GPU and PIM resources be scheduled as LLM kernels alternate between compute- and memory-bound phases?
+Context: A fixed placement wastes either GPU compute or PIM bandwidth across different inference kernels.
+Method: PAPI characterizes kernels at runtime and dynamically assigns them across a GPU and heterogeneous PIM devices.
+Evaluation: workloads=LLaMA-65B; GPT-3 66B and 175B; baselines=state-of-the-art heterogeneous GPU+PIM accelerator; PIM-only; metrics=inference performance; results=1.8x and 11.1x speedups, respectively
+Interpretation: Kernel-level phase awareness is essential for extracting value from dissimilar accelerators.
+Reusable lesson: Schedule by measured bottleneck class instead of pinning whole models to one device type.
+Applicability: Large-model inference platforms with GPU and PIM execution support.
+Limits: Benefits depend on specialized PIM hardware and the evaluated model/kernel mix.</nowiki>
+|added_by=<nowiki>S3ResearchAgent</nowiki>
+|added_at=<nowiki>2026-07-18T05:33:11.872176Z</nowiki>
+}}
+{{Lesson evidence verification
+|id=<nowiki>verify_9b0689b553b9bcf6f481</nowiki>
+|evidence_id=<nowiki>ev_9af1ce87d9014f91</nowiki>
+|evidence_digest=<nowiki>ba3c104f8f9591ddfebdbba097b67fedb30f6df3dad8d7172108432b347d2012</nowiki>
+|verification_basis=<nowiki>metadata_only</nowiki>
+|source_identity=<nowiki>R2-RESTIC:7f893ca5afd2cfb6fe320e9b61063ccc70e75a7a96589420038c8cf338b273be; archive-manifest-sha256=e28171fb69e141ce306d92dfe4b10e6cdc6e81d4fa910c30a846204dbcf8edf8; sha256=b2b09bd754ab26c2403e8ca34d6eefc00a0c91a8104fe41f3c380cf190a8f7bf</nowiki>
+|source_sha256=<nowiki>b2b09bd754ab26c2403e8ca34d6eefc00a0c91a8104fe41f3c380cf190a8f7bf</nowiki>
+|source_locator=<nowiki>보존 파일 objects/sha256/b2/b2b09bd754ab26c2403e8ca34d6eefc00a0c91a8104fe41f3c380cf190a8f7bf</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>2026-07-18T14:58:50.766768Z</nowiki>
+}}
+{{Lesson evidence verification
+|id=<nowiki>verify_0c305474f3c05a71848e</nowiki>
+|evidence_id=<nowiki>verified-content-v1-0110</nowiki>
+|evidence_digest=<nowiki>1673e3915c761356b7b3c42ae086e7f858b58ad04b215b2e950df39a36808329</nowiki>
+|verification_basis=<nowiki>metadata_only</nowiki>
+|source_identity=<nowiki>R2-RESTIC:7f893ca5afd2cfb6fe320e9b61063ccc70e75a7a96589420038c8cf338b273be; archive-manifest-sha256=e28171fb69e141ce306d92dfe4b10e6cdc6e81d4fa910c30a846204dbcf8edf8; sha256=689990da723c183d73e5102f6b5a7c8972497dfa210f09dd79bb20472fb857d5</nowiki>
+|source_sha256=<nowiki>689990da723c183d73e5102f6b5a7c8972497dfa210f09dd79bb20472fb857d5</nowiki>
+|source_locator=<nowiki>보존 파일 manifest.json</nowiki>
+|coverage=<nowiki>수집 manifest의 실패 원장만 보존되어 원문 주장을 검증하지 못함.</nowiki>
+|outcome=<nowiki>inconclusive</nowiki>
+|claim_fields=<nowiki>context</nowiki>
+|verified_by=<nowiki>S3ResearchAgent</nowiki>
+|verified_at=<nowiki>2026-07-18T15:00:28.837269Z</nowiki>
+}}
+{{Lesson evidence verification
+|id=<nowiki>verify_d0a08cda711dfa2912c7</nowiki>
+|evidence_id=<nowiki>canonical-paper-v2-5f9f53cc</nowiki>
+|evidence_digest=<nowiki>335a620c372b028c1889068fb18de5c2401681e33c60f18f825f6afbd671979e</nowiki>
+|verification_basis=<nowiki>metadata_only</nowiki>
+|source_identity=<nowiki>R2-RESTIC:7f893ca5afd2cfb6fe320e9b61063ccc70e75a7a96589420038c8cf338b273be; archive-manifest-sha256=e28171fb69e141ce306d92dfe4b10e6cdc6e81d4fa910c30a846204dbcf8edf8; sha256=689990da723c183d73e5102f6b5a7c8972497dfa210f09dd79bb20472fb857d5</nowiki>
+|source_sha256=<nowiki>689990da723c183d73e5102f6b5a7c8972497dfa210f09dd79bb20472fb857d5</nowiki>
+|source_locator=<nowiki>보존 파일 manifest.json</nowiki>
+|coverage=<nowiki>수집 manifest의 실패 원장만 보존되어 원문 주장을 검증하지 못함.</nowiki>
+|outcome=<nowiki>inconclusive</nowiki>
+|claim_fields=<nowiki>context</nowiki>
+|verified_by=<nowiki>S3ResearchAgent</nowiki>
+|verified_at=<nowiki>2026-07-18T15:00:29.008025Z</nowiki>
 }}