Update Log

Bundle Update Log

2026-10-08

Watch (morning)

• Watch: 2026-10-08 Asia/Shanghai AI infra 论文巡检。arXiv cs.AR/cs.DC new = Wed 10/7（cs.AR 18 新 + 交叉/替换；cs.DC 18 新 + 交叉/替换）。口径含 WSE/NoC/NoW/SoC/3D IC、LLM architecture/interconnect/accelerator/chiplet 与 agentic AI architecture / chip design。去重：HBF 已有 characterizing-hbf / hotcold / hbflex / dash 页，Lachesis 放置依据（寿命）不同；Terracotta 2610.06475 昨日已入库；其余新候选 grep 未命中。
• Ingest: Lachesis PDF + stub（arXiv:2610.08378）；DynaCore（2610.07443）；T-CCL（2610.07098，SC’26 Workshops）；NCCL M2N（2610.07516）。量化数字均复核自本地原文 PDF。
• Creation (papers): Lachesis（HBF 寿命 vs HBM-first 1.19–3.13×，3.3–12.2 device-years）；DynaCore（TTFT vs FIGLUT/Planaria 3.50×/2.97×，TPOT 36.55×/8.02×）；T-CCL（vs NCCL 最高 2.4×，受限预算 3.42×；vLLM 最高 1.31×）；NCCL M2N（单层最高 7.9×；DeepSeek-V3 权重同步 5.78→2.77 s）。
• Update: End-to-End Memory Data Path、LLM Distributed Training Collectives、M2N Communication、DNN Accelerator Systolic Dataflow、Prefill-Decode Divergence。
• Indexes: 手动同步 papers/index.md、raw/papers/index.md；未跑 generate_indexes.py。
• Considered not ingested: Ofan (2610.07230，胖树目的地感知交换机负载均衡，CCT 膨胀降 16–39×，下轮候选)；fbnic (2610.07644，Meta 自研多主机网卡运维，NSDI’27，偏运维)；NeMo-DCR (2610.08430，agentic RL 增量 bit-exact refit，1T 3% 变化 150 s vs 87.5 min)；TRANSIT (2610.07593，CPU DRAM 透明扩展训练显存)；ECO (2610.08373，AFD 能耗配置 BO，−40.5%)；Cascadia (2610.07219，975B MoE 跑在 11 台 AI PC)；TraceDSE (2610.07191，MLCAD’26，agent 做异构 SoC DSE)；ActTune (2610.08444，VLA 精度/DVFS)；SVRF (2610.07078)、Trail (2610.08483)、gem5 计算型 DRAM (2610.08186)、DySCo (2610.08268)、多 agent 持久记忆评测 (2610.07782) 等非本轮主线。

2026-10-07

Watch (morning)

• Watch: 2026-10-07 Asia/Shanghai AI infra 论文巡检。arXiv cs.AR/cs.DC new = Tue 10/6（cs.AR 23 条含交叉/替换；cs.DC 68 条）。口径含 WSE/NoC/NoW/SoC/3D IC、LLM architecture/interconnect/accelerator/chiplet 与 agentic AI architecture / chip design。去重：2608.24637（晶圆级光互连热调谐）、2609.39131（HBF）已有页，2609.14643（BigMoMo）log 已提；新候选 grep 均未命中。
• Ingest: VLA workload characterization PDF + stub（arXiv:2610.05062，ASPLOS’27）；HiNa-MoE（2610.05123，PACT’26）；Parallelism compute–comm trade-offs（2610.05305）；Terracotta（2610.06475，MICRO 2026 扩展版）。量化数字均复核自本地原文 PDF。
• Creation (papers): VLA Workload Characterization（Orin 降频省能 24–32%、Thor 1–4%；重叠执行精度 −56 pp/速度 3.9×/能耗 5.1×）；HiNa-MoE（FFN 最高 3.37×；端到端 decode 最高 2.09×）；Parallelism Trade-offs（prefill TP8 约 40% TTFT 在 NCCL；decode 有效 NVLink ~150 GB/s）；Terracotta（保留 >96% 定制收益；面积 0.03%、功耗 0.56%）。
• Update: Heterogeneous Inference、Prefill-Decode Divergence、Parallelism Transition Point、DRAM and Memory System。
• Indexes: 手动同步 papers/index.md、raw/papers/index.md；未跑 generate_indexes.py。
• Considered not ingested: PhaseGate (2610.04537，M4 UMA 上按 LLM 阶段准入 CPU 检索；检索吞吐 2.0×，NeurIPS workshop，与 EdgeAgent 同向)；Nexus (2610.05709，云-边 agent 执行平台，软件)；SparseCraft (2610.05037，LLM agent 闭环改 Gemmini RTL；周期少 2.1×，A3@MICRO workshop，用 agent 做设计而非为 agent 设计硬件)；MOLT (2610.05748，serving 与微调细粒度共享显存)；LearnSched (2610.06212，agent 自进化调度)；RetainZ (2610.05554，ZNS SSD 检查点放置)；RL-PDN (2610.06148，供电网络 RL 优化)；Alkaid (2610.04808)、AID (2610.04801)、CommuteProp (2610.05105)、SyclKittens (2610.04277)、HLS 修复/VHDL 基准等非本轮主线。

2026-10-06

Watch (morning)

• Watch: 2026-10-06 Asia/Shanghai AI infra 论文巡检。arXiv cs.AR/cs.DC new = Mon 10/5（cs.AR 13 新 + 4 交叉 + 3 替换；cs.DC 16 新 + 交叉/替换；Tue 10/6 列表上海 9 点仍未放出）。口径含 WSE/NoC/NoW/SoC/3D IC、LLM architecture/interconnect/accelerator/chiplet 与 agentic AI architecture / chip design。去重 MegaFlux/GPU-initiated comm/RapidMoE/HBF/ThunderEP/HAPMoE 等，新候选 grep 均未命中。
• Ingest: Divide and Conquer MCM GPU PDF + stub（arXiv:2610.03061）；RailWave（2610.03415）；AFORE（2610.03203）；EdgeAgent（2610.03394，ASPLOS’27）。量化数字均复核自本地原文 PDF。
• Creation (papers): MCM GPU Divide and Conquer（16-chiplet Torus/256 SM vs 同算力 SOTA MCM 性能 2.40×、能耗 −4.45×；Ring 16/64 chiplet 跌破单片 50%/10%）；RailWave（P50 通信 H800 2.02–5.84×、H20 1.74–4.36×；同需求 incast 延迟 +67.1%）；AFORE（吞吐 +10.1–17.6%、P95 ITL −7.1–9.5% vs 最强基线；迁移 0 暴露）；EdgeAgent（UMA 执行 1.29×、极端工具停顿 makespan 1.77×）。
• Update: Flattened Butterfly、Interconnection Network Design Space、LLM Distributed Training Collectives、Disaggregated Inference、Heterogeneous Inference。
• Indexes: 手动同步 papers/index.md、raw/papers/index.md；未跑 generate_indexes.py。
• Considered not ingested: Hardware-Native Joint Sparse-Quantization for Trillion-Scale MoE (2610.02241，SpTC 稀疏量化 grouped GEMM；B200 kernel 最高 1.65×，偏压缩算法)；FlashAttention on Blackwell Fixed-Shift Softmax (2610.02229，TLX 几何均值 +7.3%，kernel 增量)；Backside Clock Meshes for 2 nm BSPDN (2610.02401，背面时钟网格 skew −45%，物理设计)；RAPID Row-Parallel PIM in DRAM (2610.02502)；Coda coding-agent serving (2610.03088，准入调度软件)；VenusRL agentic RL (2610.03286，最高 4.24× 训练，软件框架)；ServeTwin (2610.02732，分布式 serving 模拟器)；WakeKV (2610.02713)、BCR (2610.02233)、CORE (2610.02235) KV 压缩/复用算法；CUDA→Tenstorrent Blackhole MLIR 迁移 (2610.02658)；100 MW AI 集群功耗管理 (2605.24461 替换版)。

2026-10-05

Watch (morning)

• Watch: 2026-10-05 Asia/Shanghai AI infra 论文巡检。arXiv cs.AR/cs.DC new = Fri 10/2（Mon 10/5 美东列表上海早晨尚未放出；相对 10/2 早报当时仅到 Thu 10/1）。口径含 WSE/NoC/NoW/SoC/3D IC、LLM architecture/interconnect/accelerator/chiplet 与 agentic AI architecture / chip design。去重 HBF-char/ThunderEP/HAPMoE/MoK/Purlin 等；cs.AR 交叉列表 2608.24637（晶圆级光互连热调谐）已有页。
• Ingest: MegaFlux PDF + stub（arXiv:2610.00671）；GPU-Initiated Communication Dissected（2610.01380）；RapidMoE（2610.01265，EuroSys’27）。量化数字均复核自本地原文 PDF。
• Creation (papers): MegaFlux（8×B200 前向/反向几何均值 1.45×/1.28×，峰值 2.14×/2.64×；vLLM DeepSeek-V4-Pro prefill 中位 1.13–1.26×）；GPU-Initiated Communication Dissected（最小 GPU 路径发起 0.7 µs/完成 4.0 µs；库额外最高 4.6 µs；~3000 连接 all-to-all 丢 59% NIC 消息率）；RapidMoE（decode 最高 3.5×、prefill 最高 2.1×；峰值 DRAM 240 GB vs KTransformers 385 GB）。
• Update: MegaMoE Kernel、LLM Distributed Training Collectives、Heterogeneous Inference、Network Interface and System Design。
• Indexes: 手动同步 papers/index.md、raw/papers/index.md；未跑 generate_indexes.py。
• Considered not ingested: Leto (2610.00687，训练 in-place 故障恢复；恢复快 3.6–6.5×，偏可靠性软件)；MoE-CORE (2610.01950，NPU 内存受限专家驻留/预取，与 RapidMoE 同向但对比口径不齐)；Serving a Revisable World (2610.01160，可中断 agent 的版本化执行，vLLM 控制面软件；修订后 TTFT 中位 −17.1%)；ePACT (2610.01784，能耗承诺跟踪)；ShatterQuant (2610.00207，16nm 块级混合精度脉动 Transformer 加速器，评测在 ViT/DiT 非 LLM)；EdgeDAE (2610.00311，VLA 扩散动作 FPGA-GPU)；Redundancy Meets Synergy (2610.00558，MoE 专家选择算法)；CONFERM CGRA / Catscan / ZTA-Q 等非本轮主线。

2026-10-02

Watch (morning)

• Watch: 2026-10-02 Asia/Shanghai AI infra 论文巡检。arXiv cs.AR/cs.DC new = Thu 10/1（相对 10/1 早报当时仅到 Wed 9/30）。口径含 WSE/NoC/NoW/SoC/3D IC、LLM architecture/interconnect/accelerator/chiplet 与 agentic AI architecture / chip design。去重 MoK/Janus/Purlin/SPLASH-layouts/SpecStream/SPIMOE/RR-Evict 等。
• Ingest: Characterizing HBF PDF + stub（arXiv:2609.39131）；ThunderEP（2609.40093）；HAPMoE（2609.39350）。量化数字均复核自本地原文 PDF。
• Creation (papers): Characterizing HBF（完成时间相对 HBM-only −36.1–87.0%；能耗最高 −55.8%；寿命 4.77→14.82 年）；ThunderEP（dispatch/combine vs NCCL 2.00×/1.53×；prefill 最高 1.66×、decode 最高 1.26×）；HAPMoE（e2e 训练吞吐最高 3.2×；非均匀 PP 再最高 +78%；搜索 <1 分钟）。
• Update: End-to-End Memory Data Path、Memory Hierarchy and Cache、LLM Distributed Training Collectives、Heterogeneous Inference。
• Indexes: 手动同步 papers/index.md、raw/papers/index.md；未跑 generate_indexes.py。
• Considered not ingested: SpecScale (2609.39334，TTS 投机服务；MATH 上吞吐 vs naive/FastTTS 2.18×/1.87×，相对 SpecStream/SPECTRA/DynBranch 增量偏搜索树服务软件)；Provenance-blind KV (2609.38706，共享 KV 正确性/合约，非体系结构主线)；Cascadia (2609.38697，AIPC 无控制面 serving)；HPC-for-Agents (2609.38723，测量/展望)；Vosti (2609.38981，确定性推理形式化)；NDS (2609.38454，通用近数据 strand)；Cobalt/CadenceRL/DScale/vSkipper/MEDEM/Joint MoE sim 等昨日已列理由仍适用；cs.AR STELLA CGRA / HLS pragma LLM 等非本轮主线。

2026-10-01

Watch (morning)

• Watch: 2026-10-01 Asia/Shanghai AI infra 论文巡检。arXiv cs.AR/cs.DC new = Wed 9/30（相对 9/30 早报当时仅到 Tue 9/29）。口径含 WSE/NoC/NoW/SoC/3D IC、LLM architecture/interconnect/accelerator/chiplet 与 agentic AI architecture / chip design。去重 SpecStream/SPIMOE/RR-Evict/EAServe/DynBranch/KV-SoK/HBF-Sim 等。
• Ingest: Mixture-of-Kittens PDF + stub（arXiv:2609.36070）；Janus（2609.36938）；Purlin（2609.36954）；SPLASH-layouts（2609.37626，异文于既有 HBF-SPLASH 2609.23816）。量化数字均复核自本地原文 PDF。
• Creation (papers): Mixture-of-Kittens（vs 最强公开基线最高 2.37×；512 GPU 生产 e2e 1.41×）；Janus（TTFT 最高 1.57–3.69×，均值 1.22–1.85×；关键路径 SSD I/O <6.5%）；Purlin（集体延迟最高 5.14×、带宽 4.50×；SGLang 离线均值 1.13×/最高 1.37×，在线交互最高 2.85×）；SPLASH-layouts（吞吐 1.3–1.73×；中位切换 <0.51% step；DOP KV 容量 +27–60%）。
• Update: NVSwitch、LLM Distributed Training Collectives、Memory Hierarchy and Cache、End-to-End Memory Data Path、Disaggregated Inference、FlashMoE Kernel。
• Indexes: 手动同步 papers/index.md、raw/papers/index.md；未跑 generate_indexes.py。
• Considered not ingested: Cobalt (2609.36959，专家共激活布局；相对 MoK 同日 MoE 训练增量偏 EP 放置)；CadenceRL (2609.36899，异构 RL rollout 调度)；DScale (2609.37532，block-diffusion 投机，相对 SpecStream/SPECTRA/DSpark 增量偏软件)；vSkipper (2609.37062，层跳过 serving 插件)；MEDEM (2609.37399，多引擎加速器 DSE)；MemExplorer (2604.16007 replaced，agentic NPU 异构内存综合)；Scepsy (2604.15186 replaced，agentic 工作流 GPU 分配)；ParaAnya (2609.36522，扩散并行采样缓存)；Joint MoE topology sim (2609.37828，仿真剖析)；FP64/INT8/FP4 Ozaki (2609.37693，数值仿真)；昨日已列 EfficientAgent/PackServe/CascadeEP/TopoEP/VarioPath 等仍适用。

2026-09-30

Watch (morning)

• Watch: 2026-09-30 Asia/Shanghai AI infra 论文巡检。arXiv cs.AR/cs.DC new = Tue 9/29（Wed 9/30 美东列表上海早晨尚未放出）。相对 9/29 早报已扫 Fri 9/25+Mon 9/28。口径含 WSE/NoC/NoW/SoC/3D IC、LLM architecture/interconnect/accelerator/chiplet 与 agentic AI architecture / chip design。去重 EAServe/DynBranch/KV-SoK/HBF-Sim/HeteroReason/Fancy Eviction 等。
• Ingest: SpecStream PDF + stub（arXiv:2609.33184）；SPIMOE（2609.34612，ICCAD’26）；RR-Evict（2609.32278）。量化数字均复核自本地原文 PDF。SCHEMA 标签新增 pim。
• Creation (papers): SpecStream（vs 卸荷基线 Qwen3 1.41× / InternLM2.5 1.32×；同 GPU 每 GPU 吞吐均值 +55.4%）；SPIMOE（vs A100 最高 8.35×；MoE FFN vs PIMoE 最高 3.33×）；RR-Evict（vs LRU P99 TTFT 最高 −75.4%、P99 uncached 最高 −65.7%）。
• Update: DSpark Speculative Decoding、Memory Hierarchy and Cache、End-to-End Memory Data Path、Disaggregated Inference、Fancy Eviction。
• Indexes: 手动同步 papers/index.md、raw/papers/index.md；未跑 generate_indexes.py。
• Considered not ingested: EfficientAgent (2609.33762，KV 卸荷何时划算的特征/策略研究，相对 SpecStream/RR-Evict 增量偏软件)；PackServe (2609.33224，agentic SLO 调度)；CascadeEP (2609.33252，MoE prefill 异步 EP)；Tessera RAG-KV (2609.32999，与既有 Tessera BSA 同名异文，需求驱动 RAG KV)；Hybrid Attention on NPUs (2609.32114)；MpFA Blackwell QK4V8 (2609.33135)；TopoEP (2609.35481)；VarioPath PCIe AlltoAllv (2609.34340)；Torch-PIM / PolyCIM / MorphAtt；AgentReplay；OLED-MoE；TempoKV；WavePP；Spexis；SmartNIC 建模；昨日已列 EAAC/PipeDRAM/CXL-SSD 等仍适用。

2026-09-29

Watch (morning)

• Watch: 2026-09-29 Asia/Shanghai AI infra 论文巡检。补扫 arXiv cs.AR/cs.DC 美东列表 Fri 9/25 + Mon 9/28（9/28 早报当时只到 Wed 9/24；Sat–Sun 无新增 AR 日栏）。口径含 WSE/NoC/NoW/SoC/3D IC、LLM architecture/interconnect/accelerator/chiplet 与 agentic AI architecture / chip design。去重 HBF-Sim/HeteroReason/Fancy Eviction/Flux/Co-Fabric/HotCold/Crossflow/EMA/Tessera/SPECTRA/SPLASH 等。
• Ingest: EAServe PDF + stub（arXiv:2609.31551，PACT’26）；DynBranch（2609.31047）；KV Cache Memory Wall SoK（2609.30854）。量化数字均复核自本地原文 PDF。
• Creation (papers): EAServe（goodput vs Dynamo 最高 4.3×、vs vLLM 1.7×；A100 4.91 req/s）；DynBranch（vs 最强基线延迟最高 −32%；vs 无复用底 −46–66%）；KV Cache Memory Wall（70B@128k +42 GB；crossover b=32→13.4k）。
• Update: Disaggregated Inference、Prefill-Decode Divergence、Memory Hierarchy and Cache、End-to-End Memory Data Path。
• Indexes: 手动同步 papers/index.md、raw/papers/index.md；未跑 generate_indexes.py。
• Considered not ingested: PipeDRAM (2609.30998，MICRO’26 PUD 通用，非 LLM 主线)；AI Datacenter HW Survey (2609.26829，昨日已跳)；EAAC (2609.30099)；CXL-SSD KV (2609.26828)；PatchKV (2609.26219，后缀编辑 KV 恢复偏软件)；Dynamo Fast Recovery (2609.25451)；Power-Aware PD (2609.24639)；Cross-Model Autoscaling (2609.29160)；KernelOPT/KREX；Blackwell agent 能耗剖析 (2609.29707，测量研究)；Xtrace；GRADE-RTL/AgenticSizing/Agentic-IC3；BitNet/Mamba CGLA 等既有跳过理由仍适用。

2026-09-28

Watch (morning)

• Watch: 2026-09-28 Asia/Shanghai AI infra 论文巡检。arXiv cs.AR/cs.DC 覆盖到 Wed 9/24（Fri–Sun 美东列表上海早晨尚未放出）；相对 9/25 早报已扫 Flux/Co-Fabric 与 9/24 HotCold/Crossflow/EMA/Tessera。口径含 WSE/NoC/NoW/SoC/3D IC、LLM architecture/interconnect/accelerator/chiplet 与 agentic AI architecture / chip design。
• Ingest: HBF-Sim PDF + stub（arXiv:2609.29246）；HeteroReason（2609.28717，MICRO’26）；Fancy Eviction（2609.28870）。量化数字均复核自本地原文 PDF。
• Creation (papers): HBF-Sim（媒体 15.94×；page-service −41.9%；条带化 3.94×）；HeteroReason（延迟 1.01–1.42×；能效 1.25–1.57×；回退 +4.2%）；Fancy Eviction（TTFT −19.9%；prefill +18.8%）。
• Update: End-to-End Memory Data Path、Memory Hierarchy and Cache、Disaggregated Inference、DSpark Speculative Decoding、Heterogeneous Inference。
• Indexes: 手动同步 papers/index.md、raw/papers/index.md；未跑 generate_indexes.py。
• Considered not ingested: AI Datacenter Hardware Survey (2609.26829，综述有用但今日优先仿真器/异构/生产轨迹三篇)；EAAC compiler–HW co-design (2609.30099，通用加速器栈)；Cross-Model Autoscaling (2609.29160，MaaS 调度软件)；Analytical Power-Aware PD Provisioning (2609.24639，运维配比)；Dynamo Fast Recovery (2609.25451，运行时恢复)；KernelOPT/KREX agent 核优化；GRADE-RTL/AgenticSizing/Agentic-IC3（相对 AHRR 增量弱）；BitNet/Mamba CGLA、Tetris RNB、CXL-SSD KV、SARA、MicroQonv、MVP CV SoC 等 9/25 已跳理由仍适用。

2026-09-25

Watch (morning)

• Watch: 2026-09-25 Asia/Shanghai AI infra 论文巡检。arXiv cs.AR/cs.DC/cs.NI 覆盖 Tue 9/22 及窗口内增量（相对 9/24 早报已扫 HotCold/Crossflow/EMA/Tessera）；去重 SPECTRA/SPLASH/Die Scaling/AHRR/CARDAN/Weave/COMET/DSec 等。口径含 WSE/NoC/NoW/SoC/3D IC、LLM architecture/interconnect/accelerator/chiplet 与 agentic AI architecture / chip design。
• Ingest: Flux PDF + stub（arXiv:2609.25949）；Co-Fabric（2609.25560）。量化数字均复核自本地原文 PDF。
• Creation (papers): Flux（iteration 最高 10×；峰值 NIC buffer >三个数量级）；Co-Fabric（延迟 >50%↓；带宽 2–5×；R1 +30–80%；互连成本 −80%、功耗 −5%）。
• Update: TPU v4 OCS、LLM Distributed Training Collectives、Interconnection Network Design Space、Interconnection Network Protocol Stack、NVSwitch、AI Infra Supernode。
• Indexes: 手动同步 papers/index.md、raw/papers/index.md；未跑 generate_indexes.py。
• Considered not ingested: BitNet CGLA (2609.27453，host-bound 2.52 tok/s)；Mamba CGLA (2609.27437)；Tetris RNB photonic scheduling (2609.25434，与 Flux 互补但今日增量弱)；CXL-SSD KV (2609.26828)；KV Working Set (2609.27746)；SARA (2609.26763)；MicroQonv；MVP CV SoC；quantum/DPD/fuzzing/tapeout education；9/24 已 ingest 的 HotCold/Crossflow/EMA/Tessera 不重复。

2026-09-24

Watch (morning)

• Watch: 2026-09-24 Asia/Shanghai AI infra 论文巡检。arXiv cs.AR/cs.DC submittedDate 覆盖 Tue 9/22–Wed 9/23（相对 9/23 早报已扫到 Mon 9/21）；去重 SPECTRA/SPLASH/Die Scaling/AHRR/CARDAN/Weave/COMET/DSec/MeshKV/HBFlex 等。口径含 WSE/NoC/NoW/SoC/3D IC、LLM architecture/interconnect/accelerator/chiplet 与 agentic AI architecture / chip design。
• Ingest: Hot–Cold HBM/HBF PDF + stub（arXiv:2609.25782）；Crossflow（2609.27085）；EMA（2609.27040）；Tessera（2609.25869）。量化数字均复核自本地原文 PDF。
• Creation (papers): HBF（14 ms TBT；resume +≈0.1 ms；会话 24×；−7.6 kW/8-GPU）；Crossflow（吞吐 geomean +16.2–17.4%；高负载最高 +43.4%）；EMA（吞吐最高 +52%；达 2× 容量静态的 96%）；Tessera（BSA 最高 6.79×；50-step 环 1.22–2.08×）。
• Update: End-to-End Memory Data Path（修 SPLASH sources 错位 + HotCold/EMA）、Disaggregated Inference、Prefill-Decode Divergence、GPU SIMT。
• Indexes: 手动同步 papers/index.md、raw/papers/index.md；未跑 generate_indexes.py。
• Considered not ingested: SARA SLO 资源分配 (2609.26763，排队论/部署优化偏软件)；KV Cache Working Set 容量规划 (2609.27746，在线规划方法)；GRADE-RTL / AgenticSizing / Agentic-IC3（EDA/验证 agent，相对昨 AHRR 增量弱）；Toki HBM FPGA 剖析 (2609.26551，工具框架)；DSAC near-threshold TPU (2609.26644，NTC 时钟，非 LLM 主线)；Multi-kW 3D HI Power (2609.24904，昨已跳)；nondeterminism GEMM (2609.25624，数值可复现内核)；PyTorch→NPU agent 转换 (2609.27249)。

2026-09-23

Watch (morning)

• Watch: 2026-09-23 Asia/Shanghai AI infra 论文巡检。arXiv cs.AR/cs.DC submittedDate 扫到 Mon 9/21；去重 CARDAN/Weave/COMET/DSec/MeshKV/HBFlex/Ask the Tool/Fathom 等。口径含 WSE/NoC/NoW/SoC/3D IC、LLM architecture/interconnect/accelerator/chiplet 与 agentic AI architecture / chip design。
• Ingest: SPECTRA PDF + stub（arXiv:2609.24847）；SPLASH（2609.23816）；Die Scaling GPU Scheduling（2609.24270）；AHRR（2609.21157）。量化数字均复核自本地原文 PDF。
• Creation (papers): SPECTRA（tile 最高 2.09×；系统级再最高 1.25×）；SPLASH（100 ms TPOT 下 3.5–11.4×）；Die Scaling（远端 HBM +67%；decode +14.3%）；AHRR（vs Direct RTL 2.6× geomean）。
• Update: GEMM vs GEMV、DSpark Speculative Decoding、DNN Systolic、End-to-End Memory Data Path、GPU SIMT、DSA Tradeoffs。
• Indexes: 手动同步 papers/index.md、raw/papers/index.md、concepts/index.md；未跑 generate_indexes.py。
• Considered not ingested: Multi-kW 3D HI Power Delivery (2609.24904，方法综述缺本轮量化增量)；WaveletECO (2609.23444，ECO agent/本地 9B，偏 EDA 工具链)；MoSim network contention sim (2609.23278，DT 仿真器)；QEffect CUDA Graph FP8 PP (2609.23536，runtime 契约偏软件)；Presage prefetch agent (2609.22636，预取搜索)；MiX VLM (2609.19683，9/21 已跳过)；NSP Nested Sequence Parallelism (2609.22755，训练 SP 软件)；NPU vehicle FINN (2609.24757)；AWE FP4 GEMM (2609.24519)；ScaleMPA RRT* (2609.24497)；CIM SAR skipping (2609.24288)；Measured Joules Learned Routes (2609.23085，能耗路由软件)；MCP-GRANITE (2609.24161，agent 接口评测)；XDNA FlashAttention case (2609.21264，编译案例)；Verification Reward Model (2609.22347)；Quality over Quantity Verilog data (2609.22765)。

2026-09-22

Watch (morning)

• Watch: 2026-09-22 Asia/Shanghai AI infra 论文巡检。arXiv cs.AR/cs.DC submittedDate 扫到 Sat 9/19；去重 MeshKV/HBFlex/Ask the Tool/Fathom/PipeSwift/Nested Parallel/Express-Mesh/WMHA 等近期条目。口径含 WSE/NoC/NoW/SoC/3D IC、LLM architecture/interconnect/accelerator/chiplet 与 agentic AI architecture / chip design。
• Ingest: CARDAN PDF + stub（arXiv:2609.21137）；Weave（2609.21483）；COMET（2609.21774）；DeepSeek DSec（2609.22978）。量化数字均复核自本地原文 PDF。
• Creation (papers): CARDAN（HBM read −26–50%；batch-1 1.15–1.31×）；Weave（MoE layer 2.89×；端到端 1.33×，摘要汇总）；COMET（400 Gb/s；资源 <1%；连接数 vs SDR 6×）；DSec（300 万 sandbox/日；>380K 并发；>5,000/s）。
• Update: DNN Accelerator Systolic Dataflow、LLM Distributed Training Collectives、Interconnection Network Protocol Stack、DSec Sandbox Platform。
• Indexes: 手动同步 papers/index.md、raw/papers/index.md、concepts/index.md；未跑 generate_indexes.py。
• Considered not ingested: ExoFlow (2609.21427，通用 Ray dataflow fault tolerance，未给 LLM/AI accelerator 增量)；Towards Efficient Serverless LLM Serving (2609.22358，K8s/Knative 启发式调度，缺体系结构增量)；From Deployment Hell to Stability (2609.22064，MLOps framework 经验)；Efficient Wide-Area Communication for Federated Foundation Model Training (2609.22673，联邦优化/仿真，非互连硬件)；The Cost-Efficiency of AI Inferencing (2609.22597，经济模型)；EdgeServerlessBench (2609.22330，通用 edge benchmark)；Agent Model Measurement (2609.21943，代码测量 agent 应用)；DaYu (2609.22090，DPU stream analytics，非本轮 AI/LLM workload)。9/21 已列教程、VLM 格式、纯 agent/评测与软件 runtime 等跳过理由仍适用。

2026-09-21

Watch (morning)

• Watch: 2026-09-21 Asia/Shanghai AI infra 论文巡检。cs.AR/cs.DC API 到 Thu 9/17（Fri–Mon 美东列表上海早晨尚未放出）。已 ingest 的 HBFlex/Ask the Tool/Fathom/PipeSwift/Nested Parallel/Express-Mesh/WMHA 等不重复。口径含 agentic AI architecture / chip design。补扫发现 MeshKV 2609.19207（9/16 投稿）在 9/18 早报漏列。
• Ingest: MeshKV PDF → raw/papers/MeshKV_NoC_KV_Cache_Fabric_2026.pdf + stub raw/papers/meshkv-noc-kv-cache-fabric.md（arXiv:2609.19207, 2026-09-16, cs.AR）。
• Creation (papers): MeshKV（流量 −58%；KV 利用率 2.1×；多流 1.9×）。
• Update: Torus、End-to-End Memory Data Path、Interconnection Network Design Space。
• Indexes: 手动同步 papers/index.md（+1）、raw/papers/index.md。未跑 generate_indexes.py。
• Considered not ingested: The Life of a Token 教程 (2609.19924)；SiliconBench 桌面统一内存评测 (2609.19169)；MiX VLM 微缩放格式 (2609.19683)；PixelFlow DiT serving (2609.20723)；Do AI Agents Understand Computer Architecture? / AutoTuring 评测 (2609.19387)；Rosetta 多代理建分析模型 (2609.19376)；Shared KV / Hybrid-State LMCache 正确性笔记 (2609.15021/15030)；DeepSeek-V4-Flash AMD gfx90a 工程笔记 (2609.15627)；Token Latency Fairness (2609.18112)；Xronos 边缘 CPU TP (2609.19909)；ASRB K8s 路由 (2609.20497)；Agentic Autoscaling 文本分类 (2609.14898)；9/18 已跳过 Rect3D/Locus/HCL/GPU ISA/VeriBug/GeoMesh/SSD-LLaMA 等仍适用。BusyBarn 仍无公开全文 PDF。

2026-09-18

Watch (morning)

• Watch: 2026-09-18 Asia/Shanghai AI infra 论文巡检。cs.AR/cs.DC pastweek 到 Thu 9/17（Fri 9/18 美东列表上海早晨尚未放出）。已 ingest 的 PipeSwift/Nested Parallel/Express-Mesh/WMHA/Trillion HBF/BOOST/PDD/UNISON 等不重复。口径含 agentic AI architecture / chip design。
• Ingest: HBFlex PDF → raw/papers/HBFlex_Flexible_Memory_HBF_LLM_2026.pdf + stub raw/papers/hbflex-flexible-memory-hbf-llm.md（arXiv:2609.18675, 2026-09-16, cs.AR）。
• Ingest: Ask the Tool PDF → raw/papers/Ask_Tool_Progress_Agent_KV_Serving_2026.pdf + stub raw/papers/ask-tool-progress-agent-kv-serving.md（arXiv:2609.18849, cs.DC）。
• Ingest: Fathom PDF → raw/papers/Fathom_Sparse_Decoding_Offloaded_KV_2026.pdf + stub raw/papers/fathom-sparse-decoding-offloaded-kv.md（arXiv:2609.17652, cs.LG；DC 列表）。
• Creation (papers): HBFlex（vs FA 1.58× / vs H3 3.30×）；Ask the Tool（p90 TTFT −20.7%/−20.8%）；Fathom（1M 1.67×；−18% bytes）。
• Update: End-to-End Memory Data Path（HBFlex/Fathom）、Disaggregated Inference（Ask the Tool）、CXL Tiered Memory（Fathom host 卸荷对照）。
• Indexes: 手动同步 papers/index.md（+3）、raw/papers/index.md。未跑 generate_indexes.py。
• Considered not ingested: Rect3D 3D-IC floorplan CAD (2609.18946)；Locus ZKP (2609.18846)；HCL MXFP4 (2609.18792)；GPU ISA encoding (2609.18662)；VeriBugBench/RTL Trojans/FAME/Lyapunov/WARD/REQAP/Analog pin/PLC ladder/analog EML/FairCompressAgent/BLADE；OAK/Ermes/Fluid Notarization/Vigil/AUPE/COMPASS-ABS/DiverseFT/UAV Edge/GeoMesh geo-training (2609.18388, WAN 训练非 WSE/NoC)/Zero-I/O FT/Token Latency Fairness/SSD-LLaMA consumer MoE (2609.18110, 消费级 SSD 栈)/Thunderbolt RDMA/ASPIRE speculative/vidax video JAX mesh/GroupKV dLLM/State P2P；Wed 已跳过 OptiPrime/FSNIC/SpecLens/CGRA/ScaleLUT/FINNAS/Carry-Through/DT-RAID/BusyBarn 等仍适用。

2026-09-17

Watch (morning)

• Watch: 2026-09-17 Asia/Shanghai AI infra 论文巡检。cs.AR recent 到 Wed 9/16（15 篇；Thu 9/17 美东列表上海早晨尚未放出）。cs.DC Wed 9/16 同步覆盖。已 ingest 的 Trillion MoE HBF/BOOST/PDD/UNISON/Vortex/Hopper Util/RoofLang 等不重复。口径含 agentic AI architecture / chip design。
• Ingest: PipeSwift PDF → raw/papers/PipeSwift_Pipeline_Parallel_Agentic_Serving_2026.pdf + stub raw/papers/pipeswift-pipeline-parallel-agentic-serving.md（arXiv:2609.16491, 2026-09-15, cs.DC）。
• Ingest: Nested Parallel PDF → raw/papers/Nested_Parallel_von_Neumann_Nested_BSP_2026.pdf + stub raw/papers/nested-parallel-von-neumann-nested-bsp.md（arXiv:2609.16787, Huawei 廖恒）。
• Ingest: Budgeted Express-Mesh PDF → raw/papers/Budgeted_Express_Mesh_NoC_2026.pdf + stub raw/papers/budgeted-express-mesh-noc.md（arXiv:2609.17057, cs.AR/cs.NI）。
• Ingest: WMHA PDF → raw/papers/World_Model_Hardware_Accelerator_WMHA_2026.pdf + stub raw/papers/world-model-hardware-accelerator-wmha.md（arXiv:2609.16244, cs.AR）。
• Creation (papers): PipeSwift（JCT 1.21–1.45× / 1.60–2.33× / 1.14–1.54×）；Nested Parallel（Nested BSP+UB peer）；Budgeted Express-Mesh（Tornado +50.7%）；WMHA（MSE×23；1.484×；68.4 mm²）。
• Update: Disaggregated Inference（PipeSwift JCT/PP）、Topology Optimization Variants + Torus（Express-Mesh）、UnifiedBus + AI Infra Supernode（Nested Parallel）、DNN Systolic（WMHA）。
• Indexes: 手动同步 papers/index.md（+4）、raw/papers/index.md。未跑 generate_indexes.py。
• Considered not ingested: OptiPrime HE-MPC 私有推理 (2609.16898, 出核心 WSE/NoC/LLM serving 口径)；Cognitive Admission Control agentic (列表错位，API 为 PipeSwift)；XMPIaaS / BOA ANNS / INT8 portable / SpecLens Verilog / CGRA / ScaleLUT / FINNAS / FSNIC / Carry-Through Checksum / DT-RAID / Nested BSP 同窗联邦/FL/农业 DS2；Tue 9/15 已跳过 GVA/VAMP/InplaceKV/AgentKV/FlashGPU-sim/BigMoMo/BrainScaleS/Cnuas 等仍适用。BusyBarn 仍无公开全文 PDF。

2026-09-16

Watch (morning)

• Watch: 2026-09-16 Asia/Shanghai AI infra 论文巡检。cs.AR recent 到 Tue 9/15（24 篇；Wed 9/16 美东列表上海早晨尚未放出）。已 ingest 的 Vortex/Hopper Util/RoofLang/Fengshui/Entwine/SAGE/Composable CXL 等不重复。口径含 agentic AI architecture / chip design。
• Ingest: Trillion MoE HBF PDF → raw/papers/Trillion_Param_MoE_HBF_Memory_Provisioning_2026.pdf + stub raw/papers/trillion-param-moe-hbf-memory-provisioning.md（arXiv:2609.15636, 2026-09-14, cs.AR）。
• Ingest: BOOST PDF → raw/papers/BOOST_Concurrent_Host_HBM_LLM_Inference_2026.pdf + stub raw/papers/boost-concurrent-host-hbm-llm-inference.md（arXiv:2609.13592, 2026-09-11, cs.DC/cs.AR）。
• Ingest: PDD PDF → raw/papers/PDD_Cross_Datacenter_Prefill_Decode_Disaggregation_2026.pdf + stub raw/papers/pdd-cross-datacenter-prefill-decode-disaggregation.md（arXiv:2609.13161, cs.AR/cs.DC；Tue 9/15 列表）。
• Ingest: UNISON PDF → raw/papers/UNISON_Near_Memory_Scheduler_LLM_Agents_2026.pdf + stub raw/papers/unison-near-memory-scheduler-llm-agents.md（arXiv:2609.09643, 2026-09-09, cs.AR；先前作 NMP KV 跳过，今日按 agentic 硬件口径入库）。
• Creation (papers): Trillion MoE HBF（状态层 1.4–4.0 s⁻¹；HBF×6 → 2.30 TB/s）；BOOST（TPOT +4.3% / 吞吐 +31%）；PDD（跨 DC BCR +37.5%）；UNISON（hit +0.3–23.1%、AMAT −22–51%、TTFT −58–89%；0.169 mm²）。
• Update: Disaggregated Inference（PDD+对照）、End-to-End Memory Data Path（BOOST/HBF/UNISON）、CXL Tiered Memory（BOOST/UNISON 对照）、Heterogeneous Inference（PDD）。
• Indexes: 手动同步 papers/index.md（+4）、raw/papers/index.md。未跑 generate_indexes.py。
• Considered not ingested: Grouped Value Attention (2609.13285, 算法侧 KV 表示压缩)；Dynamic HBM Repartitioning / VAMP (2609.13537, serving 运行时边界)；InplaceKVCache (2609.14507, KV 抽象格式)；AgentKV (2609.14872, cs.LG 软件淘汰)；FlashGPU-sim (2609.15311, 仿真器)；AMD Matrix Cores (2609.14845, 数值模型)；BigMoMo (2609.14643, 手机 MoE)；DVFS SLM (2609.13153)；NPU Eval v1.0 (2609.13166)；BrainScaleS chiplet NoC (2609.13563, 神经形态非 LLM)；mKernel (2609.13585)；ETCInfer 热调度 (2609.15230)；OpWeave (2609.14237)；Cnuas (2609.15889)；HBF Sucks 等先验跳过仍适用。BusyBarn 仍无公开全文 PDF。AccelForge/PATTON/py-kvcache/TinyML/HeatCache 等 9/14–15 跳过项仍适用。

2026-09-15

Watch (morning)

• Watch: 2026-09-15 Asia/Shanghai AI infra 论文巡检。cs.AR/cs.DC recent 到 Mon 9/14（Tue 9/15 美东列表上海早晨尚未放出）。已 ingest 的 Fengshui/Entwine/SAGE/Composable CXL 与 WaferTrans/HDA-MoE 等不重复。
• Ingest: Vortex PDF → raw/papers/Vortex_Extreme_Compression_LLM_Inference_2026.pdf + stub raw/papers/vortex-extreme-compression-llm-inference.md（arXiv:2609.12208, 2026-09-10, cs.AR）。
• Ingest: Hopper Utilization PDF → raw/papers/Dissecting_GPU_Utilization_LLM_Inference_Hopper_2026.pdf + stub raw/papers/dissecting-gpu-utilization-llm-inference-hopper.md（arXiv:2609.12923, 2026-09-11, cs.PF/cs.AR）。
• Ingest: RoofLang PDF → raw/papers/RoofLang_AI_Driven_LLM_Inference_Architecting_2026.pdf + stub raw/papers/rooflang-ai-driven-llm-inference-architecting.md（arXiv:2609.12551, 2026-09-11, cs.DC）。
• Creation (papers): Vortex（脉动 bi-flow VQ+稀疏；8.03×–23.7× / 5.68×–12.5×）；Hopper Utilization（GMMA fill 1.6–12.5%；SOL 92%→7.9%）；RoofLang（V4 峰值 decode 3.5–39.5×；B300 agent +6.23–50.1%）。
• Update: DNN Systolic（Vortex）、GPU SIMT（Hopper util）、FlashAttention-3（FA3 serving 利用率）、GEMM vs GEMV（fragment fill + bi-flow）、Disaggregated Inference（RoofLang+M 杠杆）、End-to-End Memory Data Path（KV 压缩/流量）。
• Indexes: 手动同步 papers/index.md（+3）、raw/papers/index.md。未跑 generate_indexes.py。
• Considered not ingested: py-kvcache (2609.11744, 外部 KV/NVMe 表征；有数字但相对既有 CXL/disagg 页增量偏 serving 连接器)；AccelForge (2609.11906, DSE 框架)；PATTON (2609.11392, PIM runtime)；TinyML multi-exit (2609.11939)；PyTorch operator profiling (2609.11938)；VLA robot factories (2609.12075)；HeatCache (2609.12449, 热调度)；ForgeMegakernel (2609.12379, codegen)；Argus (2609.12299, 测量编排)；CHERI/病理 BEACON/SNN/Ising/HLS agents/电网/三角格点/AMEND/UNISON/REACH/HBFSim 等先验跳过仍适用。BusyBarn 仍无公开全文 PDF。

2026-09-14

Ingest: bojieli《深入理解 AI Infra》

• Ingest: 李博杰开源书（Apache-2.0）bojieli/ai-infra-book。只拉 manuscripts/*.md（gh api / raw.githubusercontent.com），未 clone Git LFS。
• Creation (raw): 源 stub — 许可、仓库/PDF/站点、章节清单；不收录正文。
• Creation (entities): 深入理解 AI Infra — 从约束推导设计、五问、阅读地图。
• Creation (analyses): AI Infra Book — 前言 + 12 章摘要（优先深度 Ch.6/7/9/10）。
• Creation (concepts, 2/5 cap): Constraint-Driven AI Infra Design、AI Infra Supernode。Engram / 分层集合通信 / 多 rail 溢出写在章节页与既有概念补丁中。
• Update: Disaggregated Inference（Ch.9 PD/AF）、NVLink fabric（Ch.6–7 域/超售）、LLM Collectives（环/树、分层 AR、1024 卡）、Interconnection Design Space（环面 vs 交换）、Heterogeneous Inference（A100/H20 + CPU/GPU 专家）、CXL Tiered Memory（远端读类比）、Clos（QM9700 超售/rail）、Network-on-Wafer（OpenTallas 对照）、UB（OpenURMA μs）。
• SCHEMA: 增 tag book、methodology、supernode、datacenter。
• Indexes: 手工同步 analyses/index.md、entities/index.md、concepts/index.md、raw/articles/index.md。未跑 generate_indexes.py。未部署 Pages。
• Deferred: experiments/、calculations/ 代码树、archive/reviews、逐字章节转写；Ch.11 工具环境/API 成本、Ch.12 鹊桥/WAN 只做薄摘要。OpenTallas ROM 晶圆数字为书中分析、非实测硅。

Watch (morning)

• Watch: 2026-09-14 Asia/Shanghai AI infra 论文巡检。cs.AR/cs.DC recent 到 Fri 9/11（周末无新表；Mon 9/14 美东列表上海早晨尚未放出）。补扫 9/10–9/11（9/11 早报失败窗口）。已 ingest 的 WaferTrans/HDA-MoE/Sharing-a-Fabric/CIERA/REACT/FlexPosit/Huawei τ 等不重复。
• Ingest: Fengshui PDF → raw/papers/Fengshui_Chiplet_Ecosystem_BASIC_Codesign_2026.pdf + stub raw/papers/fengshui-chiplet-ecosystem-basic-codesign.md（arXiv:2609.10970, 2026-09-10, cs.AR）。
• Ingest: Entwine PDF → raw/papers/Entwine_Tiled_Computation_Fine_Grained_GPU_Comm_2026.pdf + stub raw/papers/entwine-tiled-computation-fine-grained-gpu-comm.md（arXiv:2609.11562, 2026-09-10, cs.DC）。
• Ingest: SAGE PDF → raw/papers/SAGE_Semantic_Aware_Geographic_Error_Recovery_AI_2026.pdf + stub raw/papers/sage-semantic-aware-geographic-error-recovery.md（arXiv:2609.10126, 2026-09-09, cs.AR）。
• Ingest: Composable CXL PDF → raw/papers/Composable_CXL_Memory_K8s_LLM_Serving_2026.pdf + stub raw/papers/composable-cxl-memory-k8s-llm-serving.md（arXiv:2609.10790, 2026-09-09, cs.DC）。
• Creation (papers): Fengshui（8 chiplet 池；能量/EDP×$ −48.5–97.8%）；Entwine（tile×SM；vs NCCL 1.232×）；SAGE（语义重放；vs 34-hop −28%/Ψ_del −30.1%）；Composable CXL（K8s DRA+DAX；TTFT 5.5–36.6×）。
• Update: Interconnection Design Space（Fengshui+SAGE）、Protocol Stack（SAGE）、LLM Collectives（Entwine）、NVLink fabric（Entwine）、CXL Tiered Memory（Composable CXL）、Disaggregated Inference（Fengshui+CXL）。
• Indexes: 手动同步 papers/index.md（+4）、raw/papers/index.md。未跑 generate_indexes.py。
• Considered not ingested: AccelForge (2609.11906, 加速器建模框架/DSE)；PATTON (2609.11392, 商品 PIM runtime，无新 fabric/NoC)；AMEND (2609.09823, GPU-PIM attention)；UNISON (2609.09643, NMP KV 调度)；REACH (2609.10861, HBM ECC)；HBFSim (2609.09800, HBF 仿真工具)；三角格点 NoC 路由 (2609.09746, 纯理论无 LLM)；电源/数据中心电网 (2609.11649)；CHERI/病理 BEACON/FlexSpIM SNN/Shift-Accumulate Attention；MoE Comp-Comm Overlap (2609.07536, 已于 9/10 跳过)；Tools-CC-Bench；Epoch diffusion MoE serving；EStream mobile NPU。BusyBarn 仍无公开全文 PDF。

2026-09-10

Public PDF links

• Fix: Quartz/generate_site.py 发布 papers/ 但不发布 raw/ 二进制，相对 raw/papers/*.pdf 在 https://lukebest.github.io/ 会 404。有已知 arXiv id 的论文页，把正文 PDF: 与 # Citations 里的本地 PDF 链接改为 https://arxiv.org/pdf/<id>（frontmatter sources: 仍保留本地 ingest 路径）。无 arXiv 的会议论文 / Hot Chips 幻灯改为纯文本路径（非公开本地路径），不编造 arXiv id。未跑 generate_indexes.py。

Watch (morning)

• Watch: 2026-09-10 Asia/Shanghai AI infra 论文巡检。cs.AR/cs.DC recent 到 Wed 9/9（Thu 9/10 美东列表上海早晨尚未放出）。昨日无增量日已扫过的 Mon 9/7 项不重复。
• Ingest: WaferTrans PDF → raw/papers/WaferTrans_IOMMU_free_Wafer_Scale_GPU_2026.pdf + stub raw/papers/wafertrans-iommu-free-wafer-scale-gpu.md（arXiv:2609.06125, 2026-09-05, cs.AR）。
• Ingest: HDA-MoE PDF → raw/papers/HDA_MoE_3D_NMP_Hybrid_Parallel_2026.pdf + stub raw/papers/hda-moe-3d-nmp-hybrid-parallel.md（arXiv:2609.08682, 2026-09-09, cs.AR）。
• Ingest: Sharing a Fabric PDF → raw/papers/Sharing_Fabric_Collective_Storage_Penalties_2026.pdf + stub raw/papers/sharing-fabric-collective-storage-penalties.md（arXiv:2609.06506, 2026-09-09, cs.DC）。
• Creation (papers): WaferTrans（IOMMU-free 片上翻译；vs Trans-FW 2.5×）；HDA-MoE（3D NMP hybrid 放置+调度；vs TP 1.1–3.4×）；Sharing a Fabric（存储×集体争用；DYAD 7.4× vs Lustre）。
• Update: Network-on-Wafer（UM 翻译面）、3D Stacking（3D NMP MoE）、LLM Collectives（fabric×存储）、Interconnection Design Space（一行）。
• Indexes: 手动同步 papers/index.md（+3）、raw/papers/index.md。未跑 generate_indexes.py。
• Considered not ingested: MoE Comp-Comm Overlap 资源管理 (2609.07536, FLUX/COMET SM residency，无新 fabric/NoC PHY)；Poseidon (2609.06086, 异构集群并行搜索)；EStream (2609.06551, 手机 NPU MoE prefill)；Interface-Aware KV NVM (2609.05764, 片上 NVM 量化接口)；Photonic chiplet HT (2609.06796, 安全威胁综述)；Flash KV for recsys (2609.07175)；Multichip Ising Pegasus (2609.07907)；Gutenberg NDP (2609.06691)；DejaVu unified-memory SoC (2609.05635)；Tools-CC-Bench (2609.08739, 基准)；MonoMoE/Budgeting Bytes 等 9/9 已跳过项仍适用。BusyBarn 仍无公开全文 PDF。

2026-09-09

Watch (morning)

• Watch: 2026-09-09 Asia/Shanghai AI infra 论文巡检。无增量。cs.AR/cs.DC recent/new 仍停在 Mon 9/7（Tue 9/8 与 Wed 9/9 美东列表上海早晨尚未放出）。昨日已 ingest 的 CIERA/REACT/FlexPosit/Huawei τ 与 BASP/CREDIT/Einsummable/Photonic Prefill/AInfer-PD/LEAP/DynaNDE/CHIPSMORE/Sync Tax 等不重复。
• Indexes: 无 papers 变更。未跑 generate_indexes.py。
• Considered not ingested: MonoMoE (2609.04244, H200 量化 MoE fused megakernel，无 NoC/fabric 增量)；Budgeting Bytes (2609.04238, 边缘存储字节/token roofline，非互连)；TreeFI/EOSQR/Proton irradiation/APEX-RBD/HCST/edge DVFS/GreenPipe/Serverless CVM/Atlas；LevelSyn EDA、systolic beamforming、Barnacle 区块链、JuPyLive、Iapetus 卫星 ViT。BusyBarn 仍无公开全文 PDF（仅 artifact/Zenodo）。先验跳过 Beacon/AXI4/survey 等仍适用。

2026-09-08

Watch (morning)

• Watch: 2026-09-08 Asia/Shanghai AI infra 论文巡检。cs.AR/cs.DC recent 列表到 Mon 9/7（Tue 9/8 美东列表上海早晨尚未放出）。已 ingest 的 BASP/CREDIT/Einsummable/Photonic Prefill/AInfer-PD/LEAP/DynaNDE/CHIPSMORE/Sync Tax 等不重复。
• Ingest: CIERA PDF → raw/papers/CIERA_Cross_Iteration_Exponent_Reuse_Allgather_2026.pdf + stub raw/papers/ciera-cross-iteration-exponent-reuse-allgather.md（arXiv:2609.04609, 2026-09-04, cs.DC）。
• Ingest: REACT PDF → raw/papers/REACT_Tuning_Collective_Patterns_Shared_AI_Clusters_2026.pdf + stub raw/papers/react-tuning-collective-patterns-shared-clusters.md（arXiv:2609.04417, 2026-09-03, cs.NI/cs.DC）。
• Ingest: FlexPosit PDF → raw/papers/FlexPosit_Tunable_Fractional_Precision_LLM_2026.pdf + stub raw/papers/flexposit-tunable-fractional-precision-llm.md（arXiv:2609.04724, 2026-09-04, cs.AR）。
• Ingest: Huawei τ PDF → raw/papers/Huawei_Tau_Chip_LogicFolding_Thermal_2026.pdf + stub raw/papers/huawei-tau-chip-logicfolding-thermal.md（arXiv:2609.04287, 2026-09-02, cs.AR）。
• Creation (papers): CIERA（无损指数复用 Allgather；OLMoE@16 3.70×/3.68×）；REACT（拥塞改写集体；+13–38% / ns-3 ~75%）；FlexPosit（Posit bit-serial；vs BitMoD 1.8×/1.2×）；Huawei τ（LogicFolding HB；NPU −66% iso-perf）。
• Update: LLM Collectives（CIERA+REACT），3D Stacking（τ/LogicFolding），DNN Systolic（FlexPosit 一行）。SCHEMA 补 collective/allreduce/distributed。
• Indexes: 手动同步 papers/index.md（+4）、raw/papers/index.md。未跑 generate_indexes.py。
• Considered not ingested: CIM Analog Softmax (2609.04266, 电路级 softmax、无 NoC/fabric 增量)；Atlas compound AI deploy (2609.04513, 编排)；KV low-rank adapt (2609.04263)；Prefix-cache×量化 divergence (2609.04748, serving 可复现)；GreenPipe/TreeFI/Proton/APEX-RBD/EOSQR/edge DVFS/BF16 AMX/progressive compression/MemGuard/量子/FL/电网柔性；Sep 7 已跳过 Para-Pipe/FlowTT/Analog Photonic Interposer/Latency-Aware Multi-Agent。BusyBarn 仍无公开全文 PDF。先验跳过 Beacon/AXI4/survey 等仍适用。

2026-09-07

Watch (morning)

• Watch: 2026-09-07 Asia/Shanghai AI infra 论文巡检。cs.AR/cs.DC recent 列表到 Fri 9/4（周一美东新稿上海早晨尚未放出；Labor Day 窗口）。已 ingest 的 Photonic Prefill/AInfer-PD/LEAP/DynaNDE/CHIPSMORE/Sync Tax 等不重复。
• Ingest: BASP PDF → raw/papers/BASP_Batch_Aware_Sequence_Parallelism_2026.pdf + stub raw/papers/basp-batch-aware-sequence-parallelism.md（arXiv:2609.03151, 2026-09-04, cs.DC）。
• Ingest: CREDIT PDF → raw/papers/CREDIT_DSMEM_Inter_CTA_Tiling_2026.pdf + stub raw/papers/credit-dsmem-inter-cta-tiling.md（arXiv:2609.01864, 2026-09-02, cs.DC）。
• Ingest: Einsummable PDF → raw/papers/Einsummable_Multi_GPU_Parallelism_2026.pdf + stub raw/papers/einsummable-multi-gpu-parallelism.md（arXiv:2609.03905, 2026-09-04, cs.DC）。
• Creation (papers): BASP（Ulysses 子组 A2A；1.17–1.32×）；CREDIT（DSMEM reduction-reuse；5090/H100 1.466×/1.318×）；Einsummable（自动 join-agg；LLaMA block 8.97 vs 13.65/14.87 ms）。
• Update: LLM Collectives（SP/Ulysses 子组 + 自动分解），NVLink fabric（训练侧关 NVLink 域），GPU SIMT（DSMEM 一行）。
• Indexes: 手动同步 papers/index.md（+3）、raw/papers/index.md。未跑 generate_indexes.py。
• Considered not ingested: Para-Pipe (2609.04168, 边缘异构 SoC 算子流水，非 LLM fabric)；FlowTT (2609.03459, DLRM TT embedding GPU kernel)；Analog Photonic Interposer (2609.03125, 视觉 sensor↔模拟加速器)；Latency-Aware Multi-Agent LLM on Heterogeneous GPUs (2609.03335, serving 编排)；Einsummable 同窗 Barnacle/JuPyLive/区块链/FL/5G 等出范围；Sep 4 已跳过 AceSpec/Atlas/NOVA/CREDIT 当时仅扫过标题——今日全文入库；Characterizing multi-tenancy (2609.00817) 仍表征文。BusyBarn 仍无公开全文 PDF。先验跳过 Beacon/AXI4/survey 2608.28048 等仍适用。

2026-09-04

Watch (morning)

• Watch: 2026-09-04 Asia/Shanghai AI infra 论文巡检。cs.AR recent/new = Thu 9/3 提交（含 cross）；已 ingest 的 LEAP/DynaNDE/CHIPSMORE/Sync Tax/FLINT/Maia/光互连/HYDRA/ReXpert/HCCL/DASH/DICE/Fovea/C2C/ThAME/3DLS/Mozart 与全部 hc2026-* 不重复。
• Ingest: Photonic Prefill PDF → raw/papers/Scaling_Inference_Prefill_High_Radix_Photonic_2026.pdf + stub raw/papers/scaling-inference-prefill-photonic.md（arXiv:2609.01821, 2026-09-01, cs.DC/cs.AR）。
• Ingest: AInfer-PD PDF → raw/papers/AInfer_PD_InPlace_Prefill_Decode_MoE_2026.pdf + stub raw/papers/ainfer-pd-inplace-prefill-decode-moe.md（arXiv:2609.00993, 2026-09-01, cs.DC）。
• Creation (papers): Photonic Prefill（3D 光子 4× SU BW / 1152 pod；高 batch 2.1–3.2×、跨 pod 2.2–4.5×）；AInfer-PD（Ant；turnstile+DeepEP 相位隔离；vs Normal −7.1–22.5%、vs SGLang −24.8–32.9%）。
• Update: Disaggregated Inference（同池 P/D 复用一行 + 光学 DES），NVLink NVSwitch Scale-Up Fabric（光学 1152 pod 对照），NVIDIA CPO Roadmap（推理 prefill 量化）。
• Indexes: 手动同步 papers/index.md（+2）、raw/papers/index.md。未跑 generate_indexes.py。
• Considered not ingested: AceSpec (2609.02514, edge-cloud WAN speculative，出 WSE/NoC/NoW)；Atlas 3DGS VR (2609.02352)；NOVA eNVM on-chip training (2609.01948)；RunSoC automotive (2609.01614)；Batch Before You Time EDA (2609.02470)；H3DNAS / FORGE MCU / GadIR / HDL repair；CREDIT DSMEM；MeanField GPU scheduling；Characterizing multi-tenancy AI training (2609.00817, 表征)；Just Talk Once split FL (2609.01457)。先验跳过仍适用：Block-Diffusion/VARA/HBQ、Beacon、AXI4、LLM-H、survey 2608.28048、Gen-TAS、SNN、MeshReduce-U、Redwood、Ankhdjet、TerraceMoE、FPGA Transformer survey。BusyBarn 仍无公开全文 PDF。

2026-09-03

Watch (morning)

• Watch: 2026-09-03 Asia/Shanghai AI infra 论文巡检。cs.AR new/recent = Wed 9/2 提交（12+cross）；Thu 9/3 美东列表上海早晨可能尚未放出。已 ingest 的 CHIPSMORE/Sync Tax/FLINT/Maia/光互连/HYDRA/ReXpert/HCCL/DASH/DICE/Fovea/C2C/ThAME/3DLS/Mozart 与全部 hc2026-* 不重复。
• Ingest: LEAP PDF → raw/papers/LEAP_IMC_NoC_LLM_Inference_2026.pdf + stub raw/papers/leap-imc-noc-llm-inference.md（arXiv:2609.00857, 2026-09-01, cs.AR；ICCAD’25 扩展）。
• Ingest: DynaNDE PDF → raw/papers/DynaNDE_Near_Data_Expert_Scheduling_2026.pdf + stub raw/papers/dynande-near-data-expert-scheduling.md（arXiv:2609.00407, 2026-09-01, cs.AR）。
• Creation (papers): LEAP（NUS；IMC+NMC+INC；LEAP-D 片上 PD；vs A100 ≥2.55×/≥71.94×，vs H100 1.52×/24.91×）；DynaNDE（IIT；NPU–NDP 分析模型调度；vs MoNDE prefill/decode 2.6×/2.2×）。
• Update: Disaggregated Inference（LEAP-D 片上行），Heterogeneous Inference（IMC/NMC/INC；NPU–NDP），Interconnection Network Design Space（IRCU INC），Collective-Capable NoC（对照 LEAP IRCU），CXL Tiered Memory（MoE CXL-NDP）。
• Indexes: 手动同步 papers/index.md（+2）、raw/papers/index.md。未跑 generate_indexes.py。
• Considered not ingested: Block-Diffusion edge (2609.01084, LPDDR/systolic 压缩，无 NoC/NoW/chiplet 互连增量)；VARA ReRAM (2609.00421, 通用激活稀疏 IMC)；HBQ (2609.00450, MICRO 量化)；SPEC CPU 2026 EPYC (2609.01527)；FPGA Transformer survey (2609.01212)；Analog-DB / FALCON / JENGA / energy-law / version-space / SILK replace。先验跳过仍适用：Beacon 2608.30932、AXI4 monitor、LLM-H、survey 2608.28048、Gen-TAS、SNN、MeshReduce-U、Redwood、Ankhdjet、TerraceMoE。BusyBarn（ISCA 2026）仍仅 artifact/Zenodo，无公开全文 PDF。3DLS 已入库不重复。

2026-09-02

Watch (morning)

• Watch: 2026-09-02 Asia/Shanghai AI infra 论文巡检。cs.AR new 页 = Tue 9/1 提交；Wed 9/2 美东列表上海早晨可能尚未放出。已 ingest 的 Sync Tax/FLINT/Maia/光互连/HYDRA/ReXpert/HCCL/DASH/DICE/Fovea/C2C/ThAME/3DLS/Mozart 与全部 hc2026-* 不重复。
• Ingest: CHIPSMORE PDF → raw/papers/CHIPSMORE_CIM_Chiplets_LLM_Inference_2026.pdf + stub raw/papers/chipsmore-cim-chiplets-llm-inference.md（arXiv:2608.30509, 2026-08-31, cs.AR）。
• Creation (papers): CHIPSMORE（NUS；RRAM-ACIM+SRAM-DCIM + IPCN in-network DMAC；分层 KV；非复制多请求层流水；vs H100 Mistral-7B INT8 最高 2.38× 吞吐 / 27× 能效）。
• Update: Disaggregated Inference（层流水共享权重一行），Heterogeneous Inference（CIM PE 异构），Interconnection Network Design Space（IPCN compute-in-interconnect）。
• Indexes: 手动同步 papers/index.md（+1）、raw/papers/index.md。未跑 generate_indexes.py。
• Considered not ingested: Beacon (2608.30932, LLM multi-agent chiplet HW-DSE，方法文、无新 fabric PHY/拓扑数，与 HYDRA 域重叠)；AXI4 transaction monitoring (2608.30435, Benini mixed-crit SoC)；LLM-based HW development hierarchical IRs (2608.30659, EDA)；Clock-gating MSP430 (2608.30954)；FABO routing (2608.30268)；adiabatic systolic (2608.30058)；SNN memory (2608.30444)；genomic storage thesis (2608.31004)；KORD (2608.30379)。先验跳过仍适用：survey 2608.28048、Gen-TAS、TerraceMoE gate-fail、VPP、CE-MoE、Blackwell CC、MeshReduce-U、edge Hydra 2608.25053、Intelligent Network WAN 2608.26453（出 WSE/NoC/NoW 范围）。BusyBarn（ISCA 2026）仍无公开全文 PDF。

2026-09-01

Watch (morning)

• Watch: 2026-09-01 Asia/Shanghai AI infra 论文巡检。无增量。cs.AR recent/new 到 Mon 8/31（5 条；Tue 9/1 美东列表上海早晨尚未放出）。已 ingest 的 Sync Tax/FLINT/Maia/光互连/HYDRA/ReXpert/HCCL/DASH/DICE/Fovea/C2C/ThAME/3DLS/Mozart 与全部 hc2026-* 不重复。
• Indexes: 无 papers 变更。未跑 generate_indexes.py。
• Considered not ingested: AI Hardware Accelerators survey (2608.28048, 综述无新一作互连数)；Gen-TAS (2608.28160, FPGA-GPP 任务分配)；Neuromorphic numerical solvers (2608.28387)；DeepSeq3 (2608.28188, EDA GNN)；RVV 1.0 HPC (2608.28097)；TerraceMoE (2608.27874, MoE 分层 A2A 代价模型，step-level gate 失败、实测未进 hierarchical 区)；VPP (2608.26523, Ascend chunked-prefill 软件流水)；CE-MoE layer reconfig (2608.28511, 减 A2A 的模型层型，非 fabric)；Blackwell CC TEE (2608.26575)；LLM energy token/request (2608.28044)；cache LAH/S4-FIFO (2608.27975)。Fri 8/28 已跳过项（MeshReduce-U/SNN/Redwood/Ankhdjet/HOLMES/LLM-EDA 等）仍跳过。BusyBarn（ISCA 2026）仍仅 artifact/IEEE，无公开全文 PDF。

2026-08-31

Watch (morning)

• Watch: 2026-08-31 Asia/Shanghai AI infra 论文巡检。cs.AR recent 到 Fri 8/28；Mon 8/31 美东列表上海早晨可能尚未放出。已 ingest 的 FLINT/Maia/光互连/HYDRA/ReXpert/HCCL/DASH/DICE/Fovea/C2C/ThAME/3DLS/Mozart/Iff 与全部 hc2026-* 不重复。
• Ingest: Synchronization Tax PDF → raw/papers/Synchronization_Tax_GPU_Scale_Up_Domains_2026.pdf + stub raw/papers/synchronization-tax-gpu-scale-up.md（arXiv:2608.22503, 2026-08-23, cs.DC）。
• Creation (papers): Synchronization Tax（Cornell；8-GPU 域集体 >50% 是 barrier 等待；增广 Hockney T=pα+qS/B+τ，B* 随域规模下降）。
• Update: NVLink fabric（带宽缩放 vs 域规模张力），LLM Collectives（墙钟含与 B 无关的 τ）。
• Indexes: 手动同步 papers/index.md（+1）、raw/papers/index.md。未跑 generate_indexes.py。
• Considered not ingested: MeshReduce-U (2608.26220, GNN/不规则 mesh NoC 编译器，非 LLM fabric)；SNN multicast (2608.26223)；Redwood (2608.26418, AI 设计加速器 EDA)；Ankhdjet (2608.26206, 三值 CiROM)；HOLMES yield (2608.26758)；LLM EDA orchestration (2608.27184)；DNA storage；SILK TOCTOU；vision generative edge。3D-IC Benchmark (2608.25155) 与 edge Hydra (2608.25053) 已于 8/28 跳过。BusyBarn 仍无公开 arXiv PDF。2603.22774 CPU slowdowns 替换版在窗口外。昨日跳过的 Simthesizer/NOVA/FlashAccel/TMR/SPICE 仍跳过。

2026-08-28

Watch (morning)

• Watch: 2026-08-28 Asia/Shanghai AI infra 论文巡检。cs.AR new/recent 到 Thu 8/27（7 篇新稿 + 替换）；Fri 8/28 列表尚未放出。已 ingest 的 Maia/光互连/HYDRA/ReXpert/HCCL/DASH/DICE/Fovea/C2C/ThAME/3DLS/Mozart/Iff 与全部 hc2026-* 不重复。2608.24637 v2 数字未变（2.7×/3.8×/3.3×），不重入库。
• Ingest: FLINT PDF → raw/papers/FLINT_HBF_LLM_Inference_2026.pdf + stub raw/papers/flint-hbf-llm-inference.md（arXiv:2608.25062, 2026-08-25）。
• Creation (papers): FLINT（HBF 基座 burst-buffer / phantom-plane / 只读 FTL；级联不是 DASH 双路径）。
• Update: DASH，OXMIQ HBF，TSV，DRAM，3D Stacking。
• Indexes: 手动同步 papers/index.md（+1）、raw/papers/index.md。未跑 generate_indexes.py。
• Considered not ingested: 3D-IC Benchmark Suite (2608.25155, UCLA Gupta，CATCH→3Dblox 物理设计测试集，无 LLM/NoC/NoW)；edge Hydra 表征 (2608.25053, AGX SoC + llama.cpp，与已入库 HYDRA 2608.19395 不是同一篇)；APT DiT 剪枝 (2608.25380)；BOOSTEDSOSA (2608.25346)；Syn2Logic (2608.25536)；Ising anneal (2608.26100)。cs.NI 近窗 5G/UAV/IoT/BGP/SlimTCP，无 LLM 互连。昨日跳过的 Simthesizer/NOVA/FlashAccel/TMR/SPICE/FIBER/HyperCut 与 2608.24637 v2 仍跳过。BusyBarn 仍无公开 PDF。

2026-08-27

Watch (morning)

• Watch: 2026-08-27 Asia/Shanghai AI infra 论文巡检。cs.AR new/recent 到 Wed 8/26（Tue 8/25 提交；API 最新 2608.24664 15:05 UTC）。Thu 8/27 列表尚无更新提交。cs.NI 近窗多为 5G/量子/IoT；MemChannel CXL pooling 与 WiCi 无线 GPU 不入库。已 ingest 的 HYDRA/ReXpert/HCCL/DASH/DICE/Fovea/C2C/ThAME/3DLS/Mozart/Iff 与全部 hc2026-* 不重复。
• Ingest: Maia 200 全文 PDF → raw/papers/Maia_200_Software_Defined_Dataflow_2026.pdf + stub raw/papers/maia-200-sdla.md（arXiv:2608.24664, 2026-08-25）。
• Ingest: 晶圆级光互连热调谐 PDF → raw/papers/Thermal_Tuning_Wafer_Scale_Optical_Interconnect_LLM_MoE_2026.pdf + stub raw/papers/wafer-scale-optical-interconnect-moe-thermal.md（arXiv:2608.24637, 2026-08-25）。
• Creation (papers): Maia 200 SDLA（归档全文，链到 HC 幻灯页，不重复峰值 TOPS）；晶圆级光互连热 stall。
• Update: HC Maia 200（加归档指针；幻灯 <1 µs 与全文 ~4 µs 不混用），Network-on-Wafer（光 interposer 近亲），LLM Collectives，Protocol Stack（ATLv2），3D Stacking。
• Indexes: 手动同步 papers/index.md（+2）、raw/papers/index.md。未跑 generate_indexes.py。
• Considered not ingested: Simthesizer (2608.24650, serving 仿真器+agent，无互连架构)；Pipeline-Native Transformers (2608.23841, CPU decode 带宽)；Elastic KV (2608.23658)；MemChannel CXL (2608.21731, cs.NI pooling transport)；WiCi (2608.24204)。昨日跳过的 NOVA/FlashAccel/TMR/SPICE/FIBER/HyperCut/VIPER/M3D SRAM/MCM GPU/SYNTLOG/Optalysys 仍跳过。BusyBarn 仍无公开 PDF。NVHBM 是新闻博客，不是论文。

2026-08-26

Hot Chips 2026 ingest

• Ingest: Day1/Day2 KEEP 幻灯 PDF → raw/papers/HC2026_*.pdf + Raw Source stub。数字只取 extract notes。未拷贝 Micron Confidential。
• Creation (papers): Rubin, MI455X, Helios UALoE, Crescent Island, Vera CPU, CS-4, MTIA 400, Maia 200, TPU 8, SN50, BlueField-4, Groq 3 LPX, Spectrum-X Multiplane, Jalapeño, Thor Ultra。教程/海报 7 页见同日上一小节。
• Update: Vera Rubin NVL72, Groq 3 LPX, Cerebras WSE, NVLink fabric, HCCL（MTIA 400 2D mesh / 1.2 TB/s SU）, CXL Tiered Memory（Diamond Rapids CXL 3.0 1LM/Flat2LM 三行）, TPU v4 OCS（链 TPU 8）。
• Schema: 公司加 microsoft / openai / broadcom / intel / sambanova；网络加 ualink / ualoe。
• Indexes: 手动同步 papers/index.md（+15）、entities/index.md、raw/papers/index.md。未跑 generate_indexes.py。
• Skip as full papers: Micron Confidential；Opticore；LUTs and Bolts；RISC-V / Canonical / Infineon；Day1 Arm AGI / IBM Z / Diamond Rapids（仅 CXL 三行）/ welcome / awards / Waymo / BosSemi / Fujitsu / Wildcat；Day2 Samsung LPDDR5X-PIM / XCENA MX1 / closing / Versal RF / Versal Premium Gen2。

Hot Chips 2026 tutorials / posters ingest

• Ingest: Handy / Samsung / SK hynix / d-Matrix / OXMIQ / NVIDIA Fusion 教程 PDF + Pistil 海报 PDF → raw/papers/HC2026_*.pdf + Raw Source stub。未拷贝、未 ingest Micron Confidential 教程。
• Creation (papers): Handy HBM 开场, zHBM, SK hynix packaging, d-Matrix Raptor, OXMIQ HBF, NVIDIA NVLink Fusion, Pistil。
• Update: 3D Stacking（SK HyB vs HBM4E；Samsung zHBM WoW+HCB）, Hybrid Bonding, 3D-Stacked AI Chip（Raptor 1-Hi logic-on-top）, DRAM, TSV, Network-on-Wafer（zHBM 不是晶圆级 NoW）, DASH（OXMIQ 对照）, NVLink fabric, Vera Rubin NVL72。
• Schema: 公司标签加 samsung / sk-hynix / d-matrix / oxmiq；技术加 hbf。未加 intel / microsoft / openai。
• Indexes: 手动同步 papers/index.md（+7）、raw/papers/index.md。未跑 generate_indexes.py。
• Skip: Micron Confidential；Opticore（photonic compute）；LUTs and Bolts（edu ring NoC）；RISC-V / Canonical / Infineon。

Watch (morning, no increment)

• Watch: 2026-08-26 Asia/Shanghai AI infra 论文巡检。cs.AR new/recent 聚焦 Tue 8/25（Wed 8/26 列表尚未出）；扩展核对 FlashAccel 替换版（2607.10186v2, replaced 2026-08-22）。已 ingest 的 HYDRA/ReXpert/HCCL/DASH 等不重复。
• No increment: 本轮无 WSE/NoC/NoW/3D/chiplet/LLM 互连合格全文。不硬凑入库。
• Considered not ingested: NOVA (2608.22613, 4F² VCT + peri-over-cell 两层 NMP，HB/TSV 仅作 DRAM 堆叠带宽，无 NoC/chiplet/NoW 拓扑增量)；FlashAccel (2607.10186v2, GPU 内 HBF 权重+KV 布局/SRAM 预取/管理软件，相对已 ingest 的 DASH 双路径 UCIe 无互连架构增量)；SYNTLOG FSM FPGA (2608.23288)；systolic PE approx (2608.22378)；NoTB RTL (2608.21962)；FPGA compression survey (2608.21657)；Optalysys photonic compute-in-transit (2608.21536)；M3D 6T SRAM 2nm (2608.22741)；VIPER PIM DSE (2608.23404)；MCM GPU cycle simulator (2608.22602)；TherMapNet (2608.21887)。昨日跳过的 TMR/SPICE/FIBER/HyperCut/FPGA NoC CAD/DTX/SLA/H100 load/HBM reliability/MAGMA/TokenPowerSandbox/HCRMap 仍跳过。BusyBarn 仍无公开 PDF。WATOS (2512.12279) 在窗口外。

2026-08-25

• Watch: 2026-08-25 Asia/Shanghai AI infra 论文巡检。cs.AR pastweek 最新到 Mon 8/24（6 篇）；Tue 8/25 列表尚未出。昨日已 ingest 的 HYDRA（2608.19395）不重复。
• No increment: 本轮无 WSE/NoC/NoW/3D/LLM 互连合格全文。不硬凑入库。
• Considered not ingested: TMR wide-link NoC router (2608.21288, Benini 组，512-bit/2-cycle/7nm TMR，偏 SEE 可靠性而非 LLM 互连拓扑)；SPICE MoE prefetch (2608.21240, PCIe expert offload 投机预取，软件编排)；FIBER (2608.19628)；HyperCut (2608.19296, tiled NoC 层间调度，唯一偏 LLM 的是 GPT-2 decode)；FPGA NoC CAD / DTX / SLA / H100 load / HBM reliability / MAGMA / TokenPowerSandbox / HCRMap 仍跳过。BusyBarn 仍无公开 PDF。

2026-08-24

• Watch: 2026-08-24 Asia/Shanghai AI infra 论文巡检。近 7 天 cs.AR（8/17–8/21）仍多 FPGA/RTL/GPU SIMT；WSE/NoC/NoW/3D/LLM 互连增量里只选出 1 篇有架构实质的全文。已 ingest 的 ReXpert/HCCL/DASH（08-21）、DICE（08-20）、Fovea/C2C-Explorer/ThAME（08-19）、Iff/3DLS/Mozart（08-18）不重复。
• Ingest: HYDRA PDF → raw/papers/HYDRA_Heterogeneous_Chiplet_DSE_Hybrid_LLM_2026.pdf + stub raw/papers/hydra-heterogeneous-chiplet-dse-hybrid-llm.md（arXiv:2608.19395, 2026-08-19）。
• Creation (papers): HYDRA。
• Update: C2C-Explorer, Disaggregated Inference, Interconnection Network Design Space, Interconnection Network Protocol Stack, Network-on-Wafer。
• Indexes: 手动同步 papers/index.md（+1）、raw/papers/index.md。
• Considered not ingested: FPGA NoC CAD (2608.17266)；DTX (2608.16953)；SLA scheduling (2608.16336)；H100 global load (2608.15764)；FIBER (2608.19628, GPU SIMT/寄存器)；HBM reliability (2608.19471)；MAGMA (2608.18366, FPGA GMM 视觉)；TokenPowerSandbox (2608.18149, 能耗 serving)；HCRMap (2607.11586, 3.5D MoE 映射，先前跳过、与 Mozart 重叠)。BusyBarn 仍无公开全文 PDF。本周合格增量只有 HYDRA，不硬凑 2–4 篇。

2026-08-21

• Watch: 2026-08-21 Asia/Shanghai AI infra 论文巡检。近 7–14 天 cs.AR 仍多 GNN/RTL/FPGA/SNN；WSE/NoC/NoW/3D/LLM 互连增量里选出 3 篇有架构实质的全文。已 ingest 的 Iff/3DLS/Mozart（08-18）、Fovea/C2C-Explorer/ThAME（08-19）、DICE（08-20）不重复。
• Ingest: ReXpert PDF → raw/papers/ReXpert_MoE_ReRAM_Near_Memory_Disaggregated_Serving_2026.pdf + stub raw/papers/rexpert-reram-nmc-disaggregated-moe.md（arXiv:2608.13962, 2026-08-14）。
• Ingest: HCCL PDF → raw/papers/HCCL_Collective_Communication_Meta_MTIA_300_2026.pdf + stub raw/papers/hccl-meta-mtia-300-collective-communication.md（arXiv:2608.00358；abs 自称 SC ‘26，未核程序册）。
• Ingest: DASH PDF → raw/papers/DASH_Dual_Path_HBF_MoE_LLM_Inference_2026.pdf + stub raw/papers/dash-dual-path-hbf-moe-inference.md（arXiv:2608.14333, 2026-08-14）。
• Creation (papers): ReXpert, HCCL, DASH。
• Update: Disaggregated Inference, LLM Distributed Training Collectives, DRAM and Memory System, TSV Physical Layer, 3D Stacking Technologies；交叉 ThAME、3DLS、C2C-Explorer、Meta RDMA。
• Schema: SCHEMA.md 公司标签加 meta。
• Indexes: 手动同步 papers/index.md（+3）、raw/papers/index.md。
• Considered not ingested: FPGA NoC CAD (2608.17266, FPL 2024 投稿窗口、FPGA 布局布线，无 LLM 架构增量)；DTX (2608.16953, 训练脉动/VLIW，无互连/3D/chiplet)；Dryas (2608.12934, Enzian ECI 跟踪引擎)；HBF Sucks / Potential Applications of HBF / Beyond Capacity 的姊妹短文 (2608.11668/13127/13868, 表征或短 CAL，架构增量已被 DASH 覆盖)；Hardware Design and Security (2608.05063, 安全)；SLA scheduling (2608.16336)；H100 global load (2608.15764)；BusyBarn 仍无公开全文 PDF（仅 artifact/IEEE）；先前跳过的 SHIFT/HCRMap/SiFAR/CLIP-3D/Chiplet-Contiguous/DeepStack/3D-Flow/DyPNet-MSC/Trivance/200mm InOx/2608.15118 仍无新全文增量。

2026-08-20

• Watch: 2026-08-20 Asia/Shanghai AI infra 论文巡检。近 7–14 天 cs.AR 多为 GNN/RTL/FPGA CAD，WSE/NoC/NoW/3D/LLM 互连增量很少。已 ingest 的 Iff/3DLS/Mozart（08-18）与 Fovea/C2C-Explorer/ThAME（08-19）不重复。
• Ingest: DICE PDF → raw/papers/DICE_Detailed_Inter_Chiplet_End_to_End_PHY_Modeling_2026.pdf + stub raw/papers/dice-detailed-inter-chiplet-end-to-end-phy-modeling.md（arXiv:2607.24221；PDF 页眉仍为 ISCA 2026 submission draft）。
• Creation (papers): DICE。
• Update: Interconnection Network Protocol Stack（chiplet PHY/FEC/PAM4）、C2C-Explorer、Network-on-Wafer、UB 物理层。
• Indexes: 手动同步 papers/index.md（+1）、raw/papers/index.md。
• Considered not ingested: Collective Communication for Distributed LLM Systems (2608.15118, 集群级 AR/RS/AG/A2A 教程，与已有 collectives 概念重叠、非 WSE/NoC/NoW/3D)；200 mm M3D InOx (2608.09508, 器件/工艺)；BusyBarn (ISCA 2026 晶圆级 LLM 映射+BALD，IEEE 付费、无公开 PDF，数字无法核)；Ouroboros / FlatAttention / ELMoE-3D / ATLAS 等 3–5 个月前工作超出 1–2 月窗口。先前跳过的 SHIFT/HCRMap/SiFAR/CLIP-3D/Chiplet-Contiguous/DeepStack/3D-Flow/DyPNet-MSC/Trivance 仍无新全文增量。

2026-08-19

• Watch: 2026-08-19 Asia/Shanghai AI infra 论文巡检。检索近 7–14 天并扩展到约 2 个月：WSE / NoC / NoW / 3D IC / LLM 加速器互连。昨日已 ingest 的 Iff/3DLS/Mozart 未重复。
• Ingest: Fovea PDF → raw/papers/Fovea_Physical_Implication_Aware_Wafer_Scale_DSE_2026.pdf + stub raw/papers/fovea-physical-implication-aware-wafer-scale-dse.md（arXiv:2608.03285, 2026-08-04）。
• Ingest: C2C-Explorer PDF → raw/papers/C2C_Explorer_Chip_to_Chip_Interconnect_LLM_2026.pdf + stub raw/papers/c2c-explorer-chip-to-chip-interconnect-llm.md（arXiv:2608.08611, DAC 2026 自称）。
• Ingest: ThAME PDF（v2）→ raw/papers/ThAME_3D_Memory_Enabled_Heterogeneous_MoE_2026.pdf + stub raw/papers/thame-3d-memory-enabled-heterogeneous-moe.md（arXiv:2607.17074；昨日仅摘要、今日全文）。
• Creation (papers): Fovea, C2C-Explorer, ThAME。
• Update: Network-on-Wafer（Fovea 可行域/DSE）, 3D Stacking Technologies（CBA+HB）, 3D-Stacked AI Chip, LLM Distributed Training Collectives, Cerebras WSE, Mozart / WoW / 3DLS 论文页交叉引用。
• Indexes: 手动同步 papers/index.md（+3）。
• Considered not ingested: SHIFT (2606.28754, 计算搬迁 vs 数据搬迁，对比「晶圆级 LLM 服务」但主贡献偏 runtime)；HCRMap (2607.11586, 3.5D MoE 映射，与 Mozart 重叠且偏调度)；SiFAR (2607.08973, 软件同步-free AllReduce)；HyNoC (FPGA VLIW NoC, LLaMA 只作负载)；CLIP-3D (2607.12788, 通用 3D-IC 热/布局, 非 LLM)；200 mm M3D InOx (2608.09508, 器件/工艺为主)；Chiplet-Contiguous Layout / locality simulator (2606.11718/11716, 多 chiplet GPU GEMM)；DeepStack / 3D-Flow / DyPNet-MSC / Trivance 仍缺相对 wiki 的新增量或全文无新实质。无增量条目不单列「无增量」，因本轮有三篇 ingest。

2026-08-18

• Watch (first-run): 2026-08-18 Asia/Shanghai 首轮 AI infra 论文巡检。检索 2025–2026（偏近 2–8 周）WSE / NoC / NoW / 3D IC / LLM 加速器。已有页未重复 ingest：FlooNoC collectives、Cerebras/WSE、Voxel、WaferLLM、MOCAP、hybrid bonding 综述。
• Considered not ingested: ThAME (arXiv:2607.17074, ESWEEK-26, 15.7×/9.8× 仅摘要级)；DeepStack (2604.04750, 与 Voxel DSE 重叠)；3D-Flow FlashAttention hybrid-bond (2602.11016)；DyPNet-MSC photonic NoW (ISPASS 2026)；Trivance AllReduce (2602.17254)；RPU (2602.18568)；CHIME (2601.19908)。
• Ingest: Iff et al. WoW NoW PDF → raw/papers/Network_Design_Wafer_Scale_WoW_Hybrid_Bonding_2026.pdf + stub raw/papers/network-design-wafer-scale-wow-hybrid-bonding.md（arXiv:2603.05266）。
• Ingest: 3DLS PDF → raw/papers/3DLS_3D_Logic_Stacked_Disaggregated_LLM_Serving_2026.pdf + stub raw/papers/3dls-3d-logic-stacked-disaggregated-llm-serving.md（arXiv:2607.01617, IEEE CAL 2026）。
• Ingest: Mozart PDF → raw/papers/Mozart_35D_Wafer_Scale_MoE_Training_2026.pdf + stub raw/papers/mozart-35d-wafer-scale-moe-training.md（arXiv:2603.07006）。
• Creation (papers): WoW Network Design, 3DLS, Mozart。
• Creation (concepts): Network-on-Wafer — 三条 WSI 物理路线 + 放置即拓扑。
• Update: 3D Stacking Technologies, 3D-Stacked AI Chip, Disaggregated Inference, Cerebras WSE, Post-Moore Architecture Frontiers, LLM Distributed Training Collectives, SCHEMA.md（加 now / network-on-wafer / wafer-on-wafer）。
• Indexes: 手动同步 concepts/index.md（+1）、papers/index.md（+3）。

2026-08-13

• Ingest: Corner To Corner Route 交互可视化 → raw/articles/corner-to-corner-route.html + 摘录 raw/articles/corner-to-corner-route.md（6×8 AIC 折叠多环、RBRG 转弯服务、相位约束最短路）。
• Creation: AIC Folded Multi-Ring NoC — floorplan 常量、微边周期表、H→V→H 合法性、对角 194 cyc / 53.7 mm 复现。
• Update: Linear and Ring Topology, Mesh and Torus Topology, Deterministic Routing and DOR, Topology Optimization Variants — 折叠多环/几何 DOR 交叉引用。
• Indexes: 手动同步 concepts/index.md（+1 条）。

2026-07-31

• Layer 1 Ingest (New Study Series): 3D NoC 研究 Phase 1（Layer 1 物理层）开篇，4 raw 学记 → raw/articles/3d-noc-study-{01-tsv-process-tech,02-monolithic-vs-tsv,03-hybrid-bonding,04-3d-mesh-baseline}.md（侧重 TSV 物理 + 三路线对比 + 商业现实 + Feero baseline）。
• Creation (papers): Katti TSV 2010 — TSV 综述原典入口；Batude Monolithic 2011 — Monolithic 综述入口；Hybrid Bonding Recent — Cu-Cu 直接键合综述；Feero 3D Mesh Stan 2008 — 3-D Mesh NoC 拓扑 baseline。
• Creation (concepts): TSV Physical Layer — TSV 工艺 + KOZ + 寄生 + 热 + 良率 五约束概念页；3D Stacking Technologies — TSV / Monolithic / Hybrid Bonding 三路线对比 + 对 3D NoC 设计含义。
• Update: 3D-Stacked AI Chip + Post-Moore Architecture Frontiers 反向链接在 #5 步补。SCHEMA.md tag taxonomy 加 3d / tsv / monolithic / hybrid-bonding / through-silicon-via / microbump / cu-cu / packaging / integration / sequential-integration（既有 chiplet / physical-layer 等保留，无重复）。
• Indexes: 手动同步 concepts/index.md（+2 条）、papers/index.md（+4 条）。

2026-07-30

• Ingest: Ali (@waterloo_intern) ‘22580: From GPT2 to Kimi3, Explained’ (2026-07-27 X 长文) → raw/articles/22580 From GPT2 to Kimi3, Explained.md（551 行，已存）。
• Creation (entities): Moonshot AI Kimi K3 — K3 模型实体 + 架构骨架 + 与 GPT-2 尺度对照。
• Creation (concepts): Linear Attention Evolution — GPT-2 → Linear Attn → DeltaNet → Gated DeltaNet → KDA 七年演化主线；Attention Residuals — AttnRes 深度方向选择性残差检索；Stable Latent MoE — K3 MoE 框架（latent-space + Quantile Balancing + 898 expert/16+2 active）。
• Creation (papers): Ali 22580 From GPT2 to Kimi3 — 论文摘要页（含全部章节结构 + 关键代码 + 数学公式）。
• Update: WaferLLM System — 新增 §“与 Kimi K3 的同构关系” + 相关概念交叉引用，frontmatter sources/updated 同步；Prefill-Decode Resource Divergence — 新增 Linear Attention Evolution 与 K3 交叉引用（KDA 改变 decode bandwidth 格局），frontmatter 同步。
• Update: SCHEMA.md tag taxonomy 实际未改动（新页 tags 全部命中既有 taxonomy：attention/moe/llm/architecture/moonshot/optimization/kernel/inference 等）。
• Indexes: 手动同步 concepts/index.md（+3 条）、entities/index.md（+1 条）、papers/index.md（+1 条）。

2026-07-24

• Ingest: 金观涛、华国凡《控制论与科学方法论》→ raw/books/控制论与科学方法论-金观涛-华国凡.mobi + 结构化摘录 raw/books/cybernetics-and-scientific-methodology.md（源：/home/luke/下载/控制论与科学方法论.mobi）。
• Creation: Cybernetics and Scientific Methodology, Black-Box Epistemology。
• Update: Architecture Paper Reading Methodology, Architecture Benchmark Methodology, Network Interface and System-Level Design, Quantitative Architecture Fundamentals — 黑箱/负反馈方法论交叉引用。

2026-07-22

• Ingest: 论文精读专项 paper-deepdive Day 1–8 + OVERVIEW → raw/articles/paper-deepdive-day-01.md … day-08.md、paper-deepdive-overview.md（源：openclawdata/.../paper-deepdive/）。
• Creation: CMP NoC Pareto Design Tradeoffs, High-Radix Clos Adaptive Routing, TPU v4 OCS Reconfigurable Fabric, NVLink NVSwitch Scale-Up Fabric, Paper Deep-Dive Map；papers：route-packets-not-wires, hoskote-5ghz-mesh-polaris, balfour-tiled-cmp-noc-tradeoffs, dally-virtual-channel-flow-control, kim-adaptive-routing-high-radix-clos, tpu-v4-optically-reconfigurable, nvidia-nvlink-hopper-blackwell。
• Update: WSE Reduce Algorithms, NoC Research Methodology and Case Studies, Virtual Channel Flow Control, Clos and Fat-Tree Topology, Adaptive Routing for NoC, Architecture Paper Reading Methodology, Post-Moore Architecture Frontiers, Near-Optimal Wafer-Scale Reduce, Interconn-Study 21d Knowledge Map, Cerebras WSE, Nvidia Vera Rubin NVL72 — 精读链与 OCS/NVLink 对照交叉引用。

2026-07-21

• Ingest: Dally & Towles 互连网络 Day 19–21 → raw/articles/interconn-study-21d-day-19.md … day-21.md（源：openclawdata/.../interconn-study-21d/day-19..21.md；21 天计划完结）。
• Creation: Network Interface and System-Level Design, NoC Research Methodology and Case Studies, Interconn-Study 21d Knowledge Map。
• Update: Interconnection Network Design Space, Interconnection Network Protocol Stack, Flow Control Fundamentals, Virtual Channel Flow Control, Mesh and Torus Topology, Architecture Paper Reading Methodology, NoC Router Pipeline Optimizations, Cerebras WSE, Arch-Study 30d Knowledge Map — NI/拥塞、论文案例、21 天地图交叉引用。

2026-07-17

• Ingest: Zotero 新下载 22 篇 PDF → raw/papers/*.pdf（上次批量 Find Available PDF / arXiv 直下）。
• Creation (papers + raw stubs): MOCAP, SuperInfer, Heterogeneous Computing Agents, pHost, Multi-Branch Self-Drafting, M5 CXL, DynaX, HyperMR, Comp Parallelism, Silent Data Corruptions, Cache-Resident LLC, PRESERVE, FlexInfer, Code-Form Planning, Mixed Precision Training, HCache, Cloud-Scale RPC, CosMoS, PANDA, Aurelia, Alibaba HPN, Meta RDMA.
• Creation (concepts): CXL Tiered Memory, Mixed Precision Training.
• Update: WaferLLM System, Prefill-Decode Divergence, Disaggregated Inference, DSpark Speculative Decoding, LLM Distributed Training Collectives, Heterogeneous Inference, SCHEMA.md（补 speculative-decoding / dataflow / cxl / rdma）。
• Indexes: generate_indexes.py 重生成。

2026-07-14

• Ingest: Goossens et al. Æthereal NoC PDF → raw/papers/Aethereal_Network_on_Chip_Concepts_Architectures_Implementations_2005.pdf（Zotero: 4PFJG7KE, IEEE MDT 2005, DOI 10.1109/MDT.2005.99）。
• Creation: Æthereal NoC, aethereal-network-on-chip.md, raw/papers/aethereal-network-on-chip.md。
• Update: Switching Principles, Flow Control Fundamentals, Virtual Channel Flow Control, NoC Router 微架构, Deterministic Execution, Cerebras Color Mechanism, Interconnection Network Design Space — TDM GS / 确定性通信对照交叉引用。

2026-07-13

• Ingest: Dally & Towles 互连网络 Day 15–18 → raw/articles/interconn-study-21d-day-15.md … day-18.md（源：openclawdata/.../interconn-study-21d/day-15..18.md）。
• Creation: Flow Control Fundamentals, Virtual Channel Flow Control, NoC Router Pipeline and Allocators, NoC Router Pipeline Optimizations。
• Update: Switching Principles, NoC Router 微架构, Deadlock-Free Routing CDG and Dally Theorem, Duato Escape VC Deadlock-Free Routing, Topology Optimization Variants, Cerebras Color Mechanism, Interconnection Network Design Space, Cerebras WSE — 流控/VC/流水线交叉引用。
• Ingest: 体系结构 30 天学习笔记 Day 27–30 → raw/articles/arch-study-30d-day-27.md … day-30.md（源：openclawdata/.../arch-study-30d/day-27..30.md）。
• Creation: LLM Distributed Training Collectives, Architecture Paper Reading Methodology, Post-Moore Architecture Frontiers, Arch-Study 30d Knowledge Map。
• Update: AllReduce Algorithms, WSE Reduce Algorithms（补 FRED/FREDR 笔记命名对照）, Parallelism Transition Point, WSE Quantitative Architecture Analysis, Cerebras WSE, Architecture Benchmark Methodology, DNN Accelerator Systolic Dataflow, Quantitative Architecture Fundamentals, Interconnection Network Design Space — Day 27–30 交叉引用。

2026-07-22

• Ingest: 4 篇 layout/NoC paper PDF → raw/papers/MAERI_*, SIGMA_*, SmartMem_*, Venus_*。
• Creation: MAERI, SIGMA, SmartMem, Venus, Layout-Aware NoC and Flexible Dataflow Accelerators。
• Update: WaferLLM Compiler Research Gaps — 新增 Gap 7 (layout-aware mesh GEMV) + 阶段 B 加 3 个新 pass。
• Update: Cerebras WSE, FEATHER Accelerator — 反向链接到新概念页。
• Validation: validate_bundle.py 通过；11 个 index 自动重生成。

2026-07-09

• Ingest: Dally & Towles 互连网络 Day 13–14 → raw/articles/interconn-study-21d-day-13.md、day-14.md（源：openclawdata/.../interconn-study-21d/day-13..14.md）。
• Creation: Deadlock-Free Routing CDG and Dally Theorem, Duato Escape VC Deadlock-Free Routing.
• Update: Adaptive Routing for NoC, Deterministic Routing and DOR, Interconnection Network Design Space, Mesh and Torus Topology, NoC Router 微架构, Switching Principles, Cerebras WSE — CDG/Dally、逃逸 VC、协议层死锁、Mesh vs Torus 交叉引用。
• Ingest: 体系结构 30 天学习笔记 Day 25–26 → raw/articles/arch-study-30d-day-25.md、day-26.md（源：openclawdata/.../arch-study-30d/day-25..26.md）。
• Creation: DNN Accelerator Systolic Dataflow, WSE Quantitative Architecture Analysis.
• Update: GPU SIMT Architecture, DSA Processor Design Tradeoffs, Eyeriss Accelerator, Plasticine Accelerator, Cerebras WSE, GEMM vs GEMV in LLM Inference, WaferLLM System — 脉动/WS·OS·RS、Amdahl/Roofline/Mesh 量化、SLA vs TPU 交叉引用。

2026-07-07

• Creation: GEMM vs GEMV in LLM Inference — 算子基础概念页：算术强度公式、Roofline、H100 decode <1% 峰值 FLOPS、prefill/decode 对应、编译器优化空间。

• Update: Prefill-Decode Resource Divergence, WaferLLM System, FlashDecoding++, WaferLLM Compiler Research Gaps — 反向链接到新 GEMM vs GEMV 页。

• Ingest: 4 篇 paper PDF → raw/papers/LoopLynx_*, SambaNova_SN40L_*, LLM_Inference_Acceleration_*, AI_Accelerators_LLM_*（arXiv: 2504.09561, 2405.07518, 2410.04466, 2506.00008）。

• Creation: LoopLynx, SambaNova SN40L, LLM Inference Hardware Survey, AI Accelerators Cross-Architecture, vLLM, WaferLLM Compiler Research Gaps。

• Update: WaferLLM System — 补 §7.5/§8 作者承认的 3 个未解瓶颈，Cerebras WSE — 新增 compiler research gaps 链接。

• Validation: validate_bundle.py 通过（4 个新 raw + 1 个新 analyses + 4 个新 paper 摘要 + 1 个新 concept）；11 个 index 自动重生成。

• Ingest: WaferLLM PDF → raw/papers/WaferLLM_LLM_Inference_at_Wafer_Scale_2025.pdf（Zotero: arXiv:2502.04563v3）。

• Creation: WaferLLM System, waferllm-wafer-scale-llm-inference.md, raw/papers/waferllm-wafer-scale-llm-inference.md.

• Update: Cerebras WSE, Prefill-Decode Resource Divergence, SpaDA Programming Language, DSA Processor Design Tradeoffs — PLMR/MeshGEMM/V、KV shift、WSE LLM serving 交叉引用。

• Ingest: 体系结构 30 天学习笔记 Day 24 → raw/articles/arch-study-30d-day-24.md（源：openclawdata/.../arch-study-30d/day-24.md）。

• Creation: GPU SIMT Architecture.

• Update: Multicore SMT and NUCA, Instruction-Level Parallelism, Branch Prediction, DRAM and Memory System, DSA Processor Design Tradeoffs — SIMT/Warp/Tensor Core、Roofline、WSE 对照交叉引用。

• Ingest: Dally & Towles 互连网络 Day 12 → raw/articles/interconn-study-21d-day-12.md（源：openclawdata/.../interconn-study-21d/day-12.md）。

• Creation: Adaptive Routing for NoC.

• Update: Deterministic Routing and DOR, Interconnection Network Design Space, NoC Router 微架构, Cerebras WSE — 最小/VRR/VC、Duato 预告、DOR 选型交叉引用。

2026-07-06

• Ingest: Dally & Towles 互连网络 21 天学习笔记 Day 9–11 → raw/articles/interconn-study-21d-day-09.md、day-10.md、day-11.md（源：openclawdata/.../interconn-study-21d/day-09..11.md）。
• Creation: Topology Optimization Variants, Deterministic Routing and DOR.
• Update: Interconnection Topology Metrics — 六拓扑统一比较、选型决策树（Day 10）；Mesh and Torus Topology, Interconnection Network Design Space, Butterfly and MIN Topology, Flattened Butterfly Topology, Collective-Capable NoC, NoC Router 微架构, Cerebras WSE — Folding/CMesh/Express、Dally 1990、XY/e-cube 路由交叉引用。
• Ingest: 体系结构 30 天学习笔记 Day 21–23 → raw/articles/arch-study-30d-day-21.md、day-22.md、day-23.md（源：openclawdata/.../arch-study-30d/day-21..23.md）。
• Creation: NoC Fundamentals (H&P Appendix F), End-to-End Memory Data Path, Multicore SMT and NUCA.
• Update: Interconnection Network Cost Model, DRAM and Memory System, Memory Hierarchy and Cache, Cache Coherence, Memory Consistency Model, SSD and NVMe Storage System, Quantitative Architecture Fundamentals, NoC Router 微架构, Cerebras WSE — 存储篇综合、NoC 五问、SMT/NUCA/Amdahl 交叉引用。
• Ingest: TileLoom PDF → raw/papers/TileLoom_Automatic_Dataflow_Planning_2026.pdf（Zotero: arXiv:2512.22168v2）。
• Creation: TileLoom Compiler, tileloom-automatic-dataflow-planning.md, raw/papers/tileloom-automatic-dataflow-planning.md.
• Update: SpaDA Programming Language, Plasticine Accelerator, Collective-Capable NoC, DSA Processor Design Tradeoffs — Triton/Helion dataflow planning vs WSE SpaDA 交叉引用。

2026-07-03

• Ingest: Constable 精读笔记 → raw/reports/constable-deepdive.md（源：openclawdata/.../superscalar-cpu/constable-deepdive.md）。
• Creation: Constable Load Elimination, constable-load-elimination.md.
• Update: Superscalar CPU Research (2023-2026), Out-of-Order Execution, Instruction-Level Parallelism, Prefill-Decode Resource Divergence — SLD/RMT/AMT、ISCA’24 Best Paper 交叉引用。
• Ingest: OpenClaw 超标量 CPU 研究综述 → raw/reports/superscalar-cpu-final-report.md、raw/reports/superscalar-cpu-report.md（源：openclawdata/.../superscalar-cpu/FINAL-report.md）。
• Creation: Superscalar CPU Research (2023-2026), superscalar-cpu-research-2023-2026.md.
• Update: Branch Prediction, Out-of-Order Execution, Instruction-Level Parallelism, DSA Processor Design Tradeoffs, Cerebras WSE, Prefill-Decode Resource Divergence, Memory Consistency Model — Constable/Bullseye/Prophet/CVA6S+ 与 WSE/LLM 交叉引用。
• Ingest: Dally & Towles 互连网络 21 天学习笔记 Day 8 → raw/articles/interconn-study-21d-day-08.md（源：openclawdata/.../interconn-study-21d/day-08.md）。
• Creation: Butterfly and MIN Topology.
• Update: Clos and Fat-Tree Topology, Switching Networks, Flattened Butterfly Topology, Mesh and Torus Topology — Butterfly/Omega/Banyan/Batcher-Banyan 与 Clos/Mesh 交叉引用。
• Ingest: 体系结构 30 天学习笔记 Day 20 → raw/articles/arch-study-30d-day-20.md（源：openclawdata/.../arch-study-30d/day-20.md）。
• Creation: SSD and NVMe Storage System.
• Update: DRAM and Memory System, Memory Hierarchy and Cache, Cerebras WSE, CMX & STX, Prefill-Decode Resource Divergence, Inference Capacity Trap — FTL/RAID/NVMe/io_uring、memoryX、KV tier 交叉引用。
• Ingest: 郑启航 知乎「分布式存储架构下的矩阵乘与编译器」→ raw/articles/分布式存储架构下的矩阵乘与编译器.md（已有 clippings；补 OKF frontmatter）。
• Creation: Distributed GEMM Algorithms, distributed-gemm-and-compiler.md.
• Update: Mesh and Torus Topology, Linear and Ring Topology, AllReduce Algorithms, Parallelism Transition Point, Graphcore IPU, SpaDA Programming Language — Cannon/SUMMA/2.5D/3D GEMM 与 T10 rTensor 交叉引用。

2026-06-24

• Ingest: Rabenseifner 2004 MPI collective reduction ICCS PDF → raw/papers/Rabenseifner_Collective_Reduction_Operations_2004.pdf（Zotero: ICCS 2004, LNCS 3036）。
• Creation: AllReduce Algorithms, rabenseifner-collective-reduction-operations.md, raw/papers/rabenseifner-collective-reduction-operations.md.
• Update: WSE Reduce Algorithms, Linear and Ring Topology, Interconnection Network Cost Model, Parallelism Transition Point, Near-Optimal Wafer-Scale Reduce — MPI Ring/RHD 与 WSE collective 谱系交叉引用。
• Ingest: Aimuyo et al. 2025 FlashMoE NeurIPS PDF → raw/papers/FlashMoE_Fast_Distributed_MoE_Single_Kernel_2025.pdf（Zotero: arXiv:2506.04667）。
• Creation: FlashMoE Kernel, flashmoe-fast-distributed-moe-single-kernel.md, raw/papers/flashmoe-fast-distributed-moe-single-kernel.md.
• Update: MegaMoE Kernel, M2N Communication, Disaggregated Inference, Parallelism Transition Point, MegaScale-Infer — MoE EP kernel 栈交叉引用。
• Ingest: Shah et al. 2024 FlashAttention-3 PDF → raw/papers/FlashAttention3_Asynchrony_Low_Precision_2024.pdf（Zotero: arXiv:2407.08608）。
• Creation: FlashAttention-3, flashattention-3-asynchrony-low-precision.md, raw/papers/flashattention-3-asynchrony-low-precision.md.
• Update: FlashAttention, FlashAttention-2, FlashDecoding++, Prefill-Decode Resource Divergence, flashattention-2-faster-attention.md, flashdecoding-plus-plus-llm-gpu-inference.md — FA→FA2→FA3 谱系补全。
• Ingest: Dao et al. 2022 FlashAttention NeurIPS PDF → raw/papers/FlashAttention_Fast_IO_Aware_Attention_2022.pdf（Zotero: arXiv:2205.14135, NeurIPS 2022）。
• Creation: FlashAttention, flashattention-io-aware-exact-attention.md, raw/papers/flashattention-io-aware-exact-attention.md.
• Update: FlashAttention-2, FlashDecoding++, Prefill-Decode Resource Divergence, flashattention-2-faster-attention.md, flashdecoding-plus-plus-llm-gpu-inference.md — FA→FA2→FlashDecoding 谱系交叉引用。
• Ingest: Dao 2023 FlashAttention-2 PDF → raw/papers/FlashAttention2_Faster_Attention_2023.pdf（Zotero: arXiv:2307.08691）。
• Creation: FlashAttention-2, flashattention-2-faster-attention.md, raw/papers/flashattention-2-faster-attention.md.
• Update: FlashDecoding++, Prefill-Decode Resource Divergence, dspark-speculative-decoding.md — prefill attention vs decode kernel 栈交叉引用。
• Ingest: Hong et al. 2024 FlashDecoding++ PDF → raw/papers/FlashDecoding_PlusPlus_LLM_Inference_GPUs_2024.pdf（Zotero: arXiv:2311.01282）。
• Creation: FlashDecoding++, flashdecoding-plus-plus-llm-gpu-inference.md, raw/papers/flashdecoding-plus-plus-llm-gpu-inference.md.
• Update: Prefill-Decode Resource Divergence, DSpark Speculative Decoding, Heterogeneous Inference, dspark-speculative-decoding.md — decode kernel vs speculative/异构推理交叉引用。
• Ingest: Prabhakar et al. 2017 Plasticine ISCA PDF → raw/papers/Plasticine_Reconfigurable_Parallel_Patterns_2017.pdf（Zotero: ISCA 2017, DOI 10.1145/3079856.3080256）。
• Creation: Plasticine Accelerator, plasticine-reconfigurable-parallel-patterns.md, raw/papers/plasticine-reconfigurable-parallel-patterns.md.
• Update: Basic Data-Flow Processor, SpaDA Programming Language, DSA Processor Design Tradeoffs, FEATHER Accelerator, Eyeriss Accelerator — parallel patterns CGRA、dataflow 谱系交叉引用。
• Ingest: Chen et al. 2017 Eyeriss JSSC PDF → raw/papers/Eyeriss_Energy_Efficient_CNN_Accelerator_2017.pdf（Zotero: JSSC 2017, DOI 10.1109/JSSC.2016.2616357）。
• Creation: Eyeriss Accelerator, eyeriss-energy-efficient-cnn-accelerator.md, raw/papers/eyeriss-energy-efficient-cnn-accelerator.md.
• Update: FEATHER Accelerator, feather-reconfigurable-accelerator.md, DSA Processor Design Tradeoffs, Collective-Capable NoC, NoC Router 微架构 — RS dataflow、GIN 组播 NoC、FEATHER 固定基线交叉引用。
• Ingest: Dennis & Misunas 1975 basic data-flow processor PDF → raw/papers/Dennis_Misunas_Basic_Data_Flow_Processor_1975.pdf（Zotero: ISCA 1975, ACM 641675.642111）。
• Creation: Basic Data-Flow Processor, dennis-misunas-basic-data-flow-processor.md, raw/papers/dennis-misunas-basic-data-flow-processor.md.
• Update: Deterministic Execution, DSA Processor Design Tradeoffs, CPU Pipeline Fundamentals, SpaDA Programming Language, Cerebras WSE — 数据流架构历史交叉引用。
• Update: Collective-Capable NoC, collective-capable-noc-ml-accelerators.md — 扩充 DCA 范式。
• Ingest: Colagrande et al. 2026 collective-capable NoC PDF → raw/papers/Collective_Capable_NoC_ML_Accelerators_2026.pdf（Zotero: MLSys 2026, arXiv:2603.26438）。
• Creation: Collective-Capable NoC, collective-capable-noc-ml-accelerators.md, raw/papers/collective-capable-noc-ml-accelerators.md.
• Update: NoC Router 微架构, Mesh and Torus Topology, WSE Reduce Algorithms, Memory Consistency Model, Cerebras WSE — FlooNoC multicast/reduction/DCA/barrier 交叉引用。
• Ingest: 体系结构 30 天学习笔记 Day 19 → raw/articles/arch-study-30d-day-19.md.
• Creation: Memory Consistency Model.
• Update: Cache Coherence, Memory Fence and Barrier, Out-of-Order Execution, Deterministic Execution, DSA Processor Design Tradeoffs, Cerebras WSE, WSE Reduce Algorithms — SC/TSO/ARM、fence、CAS/MCS 锁、WSE barrier 交叉引用。
• Ingest: Dally & Towles 互连网络 21 天学习笔记 Day 7 → raw/articles/interconn-study-21d-day-07.md.
• Creation: Clos and Fat-Tree Topology.
• Update: Mesh and Torus Topology, Linear and Ring Topology, Interconnection Topology Metrics, Interconnection Network Design Space, Interconnection Network Cost Model, Switching Networks, Multi-plane Clos Topology for AI Training — Clos 定理、Fat-Tree、间接网络交叉引用。
• Ingest: FEATHER 论文 PDF → raw/papers/FEATHER_Reconfigurable_Accelerator_Dataflow_Switching_2024.pdf（Zotero: Tong et al. 2024, arXiv:2405.13170）。
• Creation: FEATHER Accelerator, feather-reconfigurable-accelerator.md, raw/papers/feather-reconfigurable-accelerator.md.
• Update: 3D-Stacked AI Chip, DSA Processor Design Tradeoffs, SpaDA Programming Language — dataflow/layout 可重构交叉引用。
• Ingest: SpaDA 论文 PDF → raw/papers/SpaDA_Spatial_Dataflow_Architecture_Programming_Language_2026.pdf（Zotero: Gianinazzi et al. 2026, arXiv:2511.09447）。
• Creation: SpaDA Programming Language, spada-spatial-dataflow-architecture.md, raw/papers/spada-spatial-dataflow-architecture.md.
• Update: Cerebras WSE, Deterministic Execution, Cerebras Color Mechanism, WSE Reduce Algorithms, Cache Coherence — SpaDA/CSL 编程模型交叉引用。
• Ingest: Dally & Towles 互连网络 21 天学习笔记 Day 6 → raw/articles/interconn-study-21d-day-06.md.
• Creation: Mesh and Torus Topology.
• Update: Interconnection Topology Metrics, Linear and Ring Topology, Interconnection Network Design Space, Interconnection Network Cost Model — 2-D Mesh/Torus、k-ary n-cube、Dally d_opt 交叉引用。
• Ingest: 体系结构 30 天学习笔记 Day 18 → raw/articles/arch-study-30d-day-18.md.
• Creation: Cache Coherence.
• Update: Memory Hierarchy and Cache, DSA Processor Design Tradeoffs, Memory Fence and Barrier, Virtual Memory and TLB, Cerebras WSE — MESI/Snooping/Directory/False Sharing 交叉引用。
• Ingest: 体系结构 30 天学习笔记 Day 17 → raw/articles/arch-study-30d-day-17.md.
• Creation: DRAM and Memory System.
• Update: Memory Hierarchy and Cache, Cerebras WSE — DRAM/HBM/内存墙交叉引用。
• Ingest: Dally & Towles 互连网络 21 天学习笔记 Day 5 → raw/articles/interconn-study-21d-day-05.md.
• Creation: Linear and Ring Topology.
• Update: Interconnection Topology Metrics, Interconnection Network Design Space — 1-D 基线、TileLink Ring 交叉引用。
• Ingest: DSpark 论文 PDF → raw/papers/DSpark_Confidence-Scheduled_Speculative_Decoding_2026.pdf（Zotero: Cheng et al. 2026）。
• Creation: DSpark Speculative Decoding, dspark-speculative-decoding.md, raw/papers/dspark-speculative-decoding.md.
• Update: DeepSeek-V4, Prefill-Decode Resource Divergence, Heterogeneous Inference — speculative decode 交叉引用。
• Ingest: Memory Fence 深度研究报告 → raw/articles/memory-fence-hardware-2026-06-28.md（源：openclawdata/.../notes/reports/）。
• Creation: Memory Fence and Barrier.
• Update: Out-of-Order Execution, Deterministic Execution, Virtual Memory and TLB, DSA Processor Design Tradeoffs, ISA Design Principles, Cerebras WSE — fence/coherence 交叉引用。
• Ingest: Dally & Towles 互连网络 21 天学习笔记 Day 3–4 → raw/articles/interconn-study-21d-day-03.md, interconn-study-21d-day-04.md.
• Creation: Interconnection Topology Metrics, Interconnection Network Cost Model.
• Update: Interconnection Network Design Space, Cerebras WSE — 拓扑度量、延迟/B_b 模型、Mesh vs Torus 权衡。
• Ingest: 体系结构 30 天学习笔记 Day 15–16 → raw/articles/arch-study-30d-day-15.md, arch-study-30d-day-16.md.
• Creation: Virtual Memory and TLB, DSA Processor Design Tradeoffs.
• Update: Memory Hierarchy and Cache, Cerebras WSE, Deterministic Execution — TLB/核心篇总结交叉引用。
• Ingest: Hennessy & Patterson 30 天体系结构学习笔记 Day 1–14 → raw/articles/arch-study-30d-day-*.md（14 文件）。
• Creation: Quantitative Architecture Fundamentals, ISA Design Principles, Numeric Formats for AI Hardware, Architecture Benchmark Methodology, CPU Pipeline Fundamentals, Instruction-Level Parallelism, Out-of-Order Execution, Branch Prediction, Memory Hierarchy and Cache.
• Update: Cerebras WSE, Deterministic Execution, FP4 Quantization-Aware Training — 与 CPU 体系结构概念交叉引用。
• Schema: 标签 taxonomy 新增 isa, pipeline, cache, power。
• Ingest: Dally & Towles 互连网络 21 天学习笔记 Day 1–2 → raw/articles/interconn-study-21d-day-01.md, interconn-study-21d-day-02.md.
• Creation: Interconnection Network Design Space, Interconnection Network Protocol Stack.
• Update: Switching Principles — 报文/虫孔交换、历史里程碑；Cerebras WSE — Mesh 度量与虫孔选型。
• Schema: 标签 taxonomy 新增 interconnect。
• Cleanup: 删除重复的 references/raw/（OKF 转换副本）；唯一原始资料目录为 raw/。megascale-infer-2504.02263.pdf、cassini-network-aware-scheduling-2308.00852.pdf 本就位于 raw/papers/（与 references/raw/papers/ 为同内容副本），无需迁移。
• Docs: README 与 OKF skill 统一为仅使用 raw/。
• Creation: Graphcore IPU, Core Group (DRAM Access Synchronization).
• Update: 3D-Stacked AI Chip, Voxel Simulator, Voxel 3D-Stacked AI Chip LLM Inference — 交叉引用拆分页。
• Schema: 标签 taxonomy 新增 graphcore。
• Ingest: Voxel 3D-Stacked AI Chip LLM Inference from raw/papers/Exploring the efficiency of 3D-stacked AI chip architecture for LLM inference with voxel.pdf (arXiv:2604.26821).
• Creation: 3D-Stacked AI Chip, Voxel Simulator.
• Update: Prefill-Decode Resource Divergence — 3D chip prefill/decode 设计空间差异。
• Creation: Converted LLM wiki at /home/luke/wiki to OKF v0.1 bundle (54 work pages + raw sources).
• Source: Karpathy-style LLM wiki (entities, concepts, papers, summaries, analyses).
• Update: Generated interactive viz.html (74 concepts, 237 cross-links).