]> 2026-09-08T15:37:40+09:00 Lesson:characterizing mobile soc for accelerating heterogeneous llm inference 6b900bb7 3000 ko On-device LLM inference on unified-memory mobile SoCs. Limits: Results are tied to one modern Snapdragon platform and its GPU/NPU software stack. The study characterizes GPU/NPU behavior and builds HeteroInfer with heterogeneous parallelism and fast synchronization on Snapdragon 8 Gen 3. S3ResearchAgent low Venue: SOSP. Year: 2025. Peak accelerator specifications obscure synchronization, operator support, and unified-bandwidth contention in real SoCs. Verification: DOI metadata and the official SOSP listing are recorded in the preserved review ledger; claim-bearing O/I/R source coverage remains pending; confidence=low. 2026-07-16T15:05:22Z 2461238.1287269 Characterizing Mobile SoC for Accelerating Heterogeneous LLM Inference. SoC-wide bandwidth and synchronization, not isolated TOPS, determine usable heterogeneous speedup. Characterizing Mobile SoC for Accelerating Heterogeneous LLM Inference workloads=mobile LLM inference on Snapdragon 8 Gen 3; baselines=GPU-only and NPU-only; metrics=inference speed; results=1.34–6.02x speedup How should mobile GPU and NPU execution be combined under a shared memory-bandwidth limit? lab Profile shared bottlenecks and partition operators across accelerators jointly. Draft 2026-07-18T14:58:23Z 2461240.1238773 2026-07-18T14:58:23Z 2461240.1238773 characterizing mobile soc for accelerating heterogeneous llm inference 6b900bb7 3000 ev_c237ba2fd2eb47e8 2026-07-16T15:05:23Z 2461238.1287384 S3ResearchAgent Characterizing Mobile SoC for Accelerating Heterogeneous LLM Inference. paper Bibliographic paper record. not_recorded characterizing mobile soc for accelerating heterogeneous llm inference 6b900bb7# 90265fbd75ff2cab56f9061ca890df1d 3000 verified-content-v1-0145 2026-07-16T18:58:12Z 2461238.2904167 S3ResearchAgent 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. paper 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. not_recorded characterizing mobile soc for accelerating heterogeneous llm inference 6b900bb7# 3327b6fde14157e2cd0fdd9037c643eb 3000 canonical-paper-v2-6b900bb7 2026-07-18T05:13:54Z 2461239.7179861 S3ResearchAgent 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. paper 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. full_text characterizing mobile soc for accelerating heterogeneous llm inference 6b900bb7# d16687664e9dbb93a5ab4f9a09487dd9 3000 2026-07-18T14:58:21Z 2461240.1238542 S3ResearchAgent verify_afa395660b05dcbe2a7b metadata_only context 보존 객체는 cookie/landing page이므로 서지 위치만 확인했고 본문 주장을 검증하지 못함. inconclusive 3799fbab6b557220cb0421a3d048d58133557c4cec9cc55394edf17b97708ddf R2-RESTIC:7f893ca5afd2cfb6fe320e9b61063ccc70e75a7a96589420038c8cf338b273be; archive-manifest-sha256=e28171fb69e141ce306d92dfe4b10e6cdc6e81d4fa910c30a846204dbcf8edf8; sha256=3799fbab6b557220cb0421a3d048d58133557c4cec9cc55394edf17b97708ddf 보존 파일 objects/sha256/37/3799fbab6b557220cb0421a3d048d58133557c4cec9cc55394edf17b97708ddf ev_c237ba2fd2eb47e8 8178f79c181e2f903b7afff9987b900367b5e1cfb8e6d96d18763b1bf4b87e20 characterizing mobile soc for accelerating heterogeneous llm inference 6b900bb7# e9159422760e0109bf88d5f104a1a145 3000 2026-07-18T14:58:22Z 2461240.1238657 S3ResearchAgent verify_1e7879fa0c68413f65b7 metadata_only context 보존 객체는 cookie/landing page이므로 서지 위치만 확인했고 본문 주장을 검증하지 못함. inconclusive 3799fbab6b557220cb0421a3d048d58133557c4cec9cc55394edf17b97708ddf R2-RESTIC:7f893ca5afd2cfb6fe320e9b61063ccc70e75a7a96589420038c8cf338b273be; archive-manifest-sha256=e28171fb69e141ce306d92dfe4b10e6cdc6e81d4fa910c30a846204dbcf8edf8; sha256=3799fbab6b557220cb0421a3d048d58133557c4cec9cc55394edf17b97708ddf 보존 파일 objects/sha256/37/3799fbab6b557220cb0421a3d048d58133557c4cec9cc55394edf17b97708ddf verified-content-v1-0145 32e1123fc0cccf1322af92ef9cdfcdea0368876a65f11cceaf402d4058ddf61b characterizing mobile soc for accelerating heterogeneous llm inference 6b900bb7# d19823c6742ea3b1892761757c396597 3000 2026-07-18T14:58:22Z 2461240.1238657 S3ResearchAgent verify_3a73bba09b397bc095b2 metadata_only context 보존 객체는 cookie/landing page이므로 서지 위치만 확인했고 본문 주장을 검증하지 못함. inconclusive 3799fbab6b557220cb0421a3d048d58133557c4cec9cc55394edf17b97708ddf R2-RESTIC:7f893ca5afd2cfb6fe320e9b61063ccc70e75a7a96589420038c8cf338b273be; archive-manifest-sha256=e28171fb69e141ce306d92dfe4b10e6cdc6e81d4fa910c30a846204dbcf8edf8; sha256=3799fbab6b557220cb0421a3d048d58133557c4cec9cc55394edf17b97708ddf 보존 파일 objects/sha256/37/3799fbab6b557220cb0421a3d048d58133557c4cec9cc55394edf17b97708ddf canonical-paper-v2-6b900bb7 587b8c417dc33053b9e977bca53c654ae2809ff2e30ef92a31f698a62ad93328 characterizing mobile soc for accelerating heterogeneous llm inference 6b900bb7# ebc1a4b604988f76412218d126542bfa 3000 2026-07-18T14:58:22Z 2461240.1238657 S3ResearchAgent verify_e928b9a88016b595911e metadata_only observation,interpretation,reusable_lesson claim-bearing O/I/R coverage was not established; confidence forced to low inconclusive 3799fbab6b557220cb0421a3d048d58133557c4cec9cc55394edf17b97708ddf R2-RESTIC:7f893ca5afd2cfb6fe320e9b61063ccc70e75a7a96589420038c8cf338b273be; archive-manifest-sha256=e28171fb69e141ce306d92dfe4b10e6cdc6e81d4fa910c30a846204dbcf8edf8; sha256=3799fbab6b557220cb0421a3d048d58133557c4cec9cc55394edf17b97708ddf; high-confidence adjudication ledger Saved HTML inspected in full; it is a cookie/landing/generic metadata page and contains no claim-bearing text covering O/I/R. canonical-paper-v2-6b900bb7 587b8c417dc33053b9e977bca53c654ae2809ff2e30ef92a31f698a62ad93328 characterizing mobile soc for accelerating heterogeneous llm inference 6b900bb7# 35503d5b1fa061e5cb4baf40f8718208 3000 2026-07-18T14:58:23Z 2461240.1238773 S3ResearchAgent verify_1fcf67e5e8e6a37158ff metadata_only context CD-001 curated source-bound correction; primary locator recorded in review. inconclusive a140f3572e510459b9eed5bf42d4e2096c4684fd4f1ee746960d6675661d1077 R2-RESTIC:7f893ca5afd2cfb6fe320e9b61063ccc70e75a7a96589420038c8cf338b273be; archive-manifest-sha256=e28171fb69e141ce306d92dfe4b10e6cdc6e81d4fa910c30a846204dbcf8edf8; sha256=a140f3572e510459b9eed5bf42d4e2096c4684fd4f1ee746960d6675661d1077; adjudicated correction content-debt-review CD-001 canonical-paper-v2-6b900bb7 587b8c417dc33053b9e977bca53c654ae2809ff2e30ef92a31f698a62ad93328 characterizing mobile soc for accelerating heterogeneous llm inference 6b900bb7# 7ca51cb67420a09191ad95c38fd30501