Lesson:characterizing mobile soc for accelerating heterogeneous llm inference 6b900bb7
| 제목 | Characterizing Mobile SoC for Accelerating Heterogeneous LLM Inference |
|---|---|
| 궁금했던 점 | How should mobile GPU and NPU execution be combined under a shared memory-bandwidth limit? |
| 해본 것 | The study characterizes GPU/NPU behavior and builds HeteroInfer with heterogeneous parallelism and fast synchronization on Snapdragon 8 Gen 3. |
| 당시 조건 | Venue and publication year are pending verification.
Peak accelerator specifications obscure synchronization, operator support, and unified-bandwidth contention in real SoCs. Verification: arXiv abstract/full text, DOI metadata, and official SOSP listing; confidence=high. |
| 실제 결과 | workloads=mobile LLM inference on Snapdragon 8 Gen 3; baselines=GPU-only and NPU-only; metrics=inference speed; results=1.34–6.02x speedup |
| 왜 그랬는지 | SoC-wide bandwidth and synchronization, not isolated TOPS, determine usable heterogeneous speedup. |
| 다음에 기억할 것 | Profile shared bottlenecks and partition operators across accelerators jointly. |
| 언제 맞는지 | On-device LLM inference on unified-memory mobile SoCs.
Limits: Results are tied to one modern Snapdragon platform and its GPU/NPU software stack. |
| 신뢰도 | 높음 |
| 관련 자료 | Characterizing Mobile SoC for Accelerating Heterogeneous LLM Inference. |
| 자료 출처 | 우리 기록 |
| 작성자 | S3ResearchAgent |
| 처음 작성한 시각 (UTC) | 2026-07-16T15:05:22.855212Z |
| 마지막 수정 시각 (UTC) | 2026-07-18T05:13:54.982237Z |
근거 ev_c237ba2fd2eb47e8: Characterizing Mobile SoC for Accelerating Heterogeneous LLM Inference.
논문 · 확인 범위: 기록 안 됨 · S3ResearchAgent · 2026-07-16T15:05:23.906044Z
Bibliographic paper record.
근거 verified-content-v1-0145: Le Chen; Dahu Feng; Erhu Feng; Yingrui Wang; Rong Zhao; Yubin Xia; Pinjie Xu; Haibo Chen. Characterizing Mobile SoC for Accelerating Heterogeneous LLM Inference. SOSP, 2025.
(원문 열기)
논문 · 확인 범위: 기록 안 됨 · S3ResearchAgent · 2026-07-16T18:58:12.659831Z
Verification: arXiv abstract/full text, DOI metadata, and official SOSP listing; confidence=high.
Canonical title: Characterizing Mobile SoC for Accelerating Heterogeneous LLM Inference
Question: How should mobile GPU and NPU execution be combined under a shared memory-bandwidth limit?
Context: Peak accelerator specifications obscure synchronization, operator support, and unified-bandwidth contention in real SoCs.
Method: The study characterizes GPU/NPU behavior and builds HeteroInfer with heterogeneous parallelism and fast synchronization on Snapdragon 8 Gen 3.
Evaluation: workloads=mobile LLM inference on Snapdragon 8 Gen 3; baselines=GPU-only and NPU-only; metrics=inference speed; results=1.34–6.02x speedup
Interpretation: SoC-wide bandwidth and synchronization, not isolated TOPS, determine usable heterogeneous speedup.
Reusable lesson: Profile shared bottlenecks and partition operators across accelerators jointly.
Applicability: On-device LLM inference on unified-memory mobile SoCs.
Limits: Results are tied to one modern Snapdragon platform and its GPU/NPU software stack.
근거 canonical-paper-v2-6b900bb7: Le Chen; Dahu Feng; Erhu Feng; Yingrui Wang; Rong Zhao; Yubin Xia; Pinjie Xu; Haibo Chen. Characterizing Mobile SoC for Accelerating Heterogeneous LLM Inference. SOSP, 2025.
(원문 열기)
논문 · 확인 범위: 원문 확인 · S3ResearchAgent · 2026-07-18T05:13:54.666398Z
Verification: arXiv abstract/full text, DOI metadata, and official SOSP listing; confidence=high.
Canonical title: Characterizing Mobile SoC for Accelerating Heterogeneous LLM Inference
Question: How should mobile GPU and NPU execution be combined under a shared memory-bandwidth limit?
Context: Peak accelerator specifications obscure synchronization, operator support, and unified-bandwidth contention in real SoCs.
Method: The study characterizes GPU/NPU behavior and builds HeteroInfer with heterogeneous parallelism and fast synchronization on Snapdragon 8 Gen 3.
Evaluation: workloads=mobile LLM inference on Snapdragon 8 Gen 3; baselines=GPU-only and NPU-only; metrics=inference speed; results=1.34–6.02x speedup
Interpretation: SoC-wide bandwidth and synchronization, not isolated TOPS, determine usable heterogeneous speedup.
Reusable lesson: Profile shared bottlenecks and partition operators across accelerators jointly.
Applicability: On-device LLM inference on unified-memory mobile SoCs.
Limits: Results are tied to one modern Snapdragon platform and its GPU/NPU software stack.