Efficiency Moves From Models Into Systems 效率优化从模型延伸至系统

This week's work attacks cost at several layers: measured 65 nm edge-AI chips, one-pass EDA optimization, sparse expert designs, and scheduling for constrained inference backends. The strongest results pair algorithmic efficiency with implementation details that expose where latency, memory, and robustness are actually won. 本周研究从多个层面压低成本：实测65 nm边缘AI芯片、单次前向的EDA优化、稀疏专家结构，以及面向受限推理后端的调度。最有价值的成果把算法效率与实现细节结合起来，明确展示延迟、内存和鲁棒性究竟从何处获得改善。

Look-back window: 7 days · 8 paper(s) 回溯窗口: 7天 · 8篇

Circuits, Architecture & Reliability 电路、架构与可靠性

A 65 nm Hypoglycemia Forecasting Engine at 11.3 nJ per Inference 单次推理11.3 nJ的65 nm低血糖预测芯片

B. Cheng, J. Liu, P. Ren, et al.

arXiv:2606.07455 · 2026-06-05T17:01:12Z

The authors fabricated a 65 nm probabilistic decision-tree engine that performs 30-minute hypoglycemia forecasting at 11.3 nJ per inference and reports an F1 score of 0.825. A hybrid exact-and-sampling architecture supports trees up to depth 12, while an on-chip RISC-V core manages a reconfigurable node array. Reported robustness to sensor noise and missing data improves by 4.1x to 16.1x over conventional tree baselines, making reliability rather than raw throughput the central contribution. 研究团队采用65 nm工艺流片了一款概率决策树引擎，用于提前30分钟预测低血糖，单次推理能耗为11.3 nJ，F1分数达到0.825。芯片以精确计算与采样混合架构支持最大深度12的树，并由片上RISC-V核心管理可重构节点阵列。相较传统决策树基线，其对传感器噪声和数据缺失的鲁棒性提高4.1至16.1倍，核心贡献更偏向可靠性而非单纯吞吐率。

A 65 nm Bayesian Compute-in-Memory Engine for Skin-Lesion Screening 面向皮肤病变筛查的65 nm贝叶斯存算一体引擎

S. Davis, L. Pei, J. Liu, et al.

arXiv:2606.07439 · 2026-06-05T16:38:02Z

This 65 nm compute-in-memory chip performs multimodal Bayesian inference for fully on-device skin-lesion screening. Its calibration-free Gaussian random-number generator reaches 16.3 fJ per sample and 168.6 GSa/s/mm², while mixture-of-Gaussian sampling improves equal-risk coverage by 1.4x and process-variation resilience by 5.5x. The result is notable for treating uncertainty generation and hardware variation as first-class circuit problems rather than software post-processing. 这款65 nm存算一体芯片执行多模态贝叶斯推理，用于完全端侧的皮肤病变筛查。其免校准高斯随机数发生器达到16.3 fJ/样本和168.6 GSa/s/mm²；高斯混合采样使等风险覆盖率提升1.4倍、工艺波动鲁棒性提升5.5倍。该工作将不确定性生成与硬件波动视为核心电路问题，而非依赖软件后处理，这是其主要价值。

EDA & Design Automation EDA与设计自动化

One-Pass Neural Optimization for Pre-Layout Signal Integrity 用于布局前信号完整性的单次神经优化

J. Withöft, W. John, E. Ecik, et al.

arXiv:2606.07463 · 2026-06-05T17:13:44Z

The paper proposes amortized neural optimization for pre-layout signal-integrity exploration, replacing repeated black-box searches with a learned policy that emits near-optimal design parameters in one deterministic forward pass. The policy is trained from analytical gradients through a differentiable surrogate and evaluated on three complex SI scenarios. The approach could make multi-corner sweeps substantially cheaper, although its practical value depends on how well the surrogate and policy generalize beyond their training space. 论文提出用于布局前信号完整性探索的摊销神经优化，以学习到的策略替代反复黑盒搜索，并通过一次确定性前向计算给出近似最优设计参数。该策略利用可微代理模型的解析梯度训练，并在三个复杂SI场景中验证。它有望显著降低多工况扫描成本，但实际价值仍取决于代理模型与策略能否泛化到训练空间之外。

AI Accelerators & Compute-in-Memory AI加速器与存算一体

Training a 120B Sparse MoE on One Eight-GPU Node 在单个八GPU节点训练120B稀疏MoE

R. Shravan

arXiv:2606.07404 · 2026-06-05T15:48:42Z

LightningLM 0.1V grows a dense seed through 5B and 9B stages into a 120B sparse MoE with 460 routed experts and top-12 routing, all on a single eight-GPU node. Only 5.93B of 118.67B stored parameters are active, and reversible recurrence reconstructs activations during backpropagation to control memory. This is explicitly a systems experience report, so its value lies in reproducible scaling disciplines and failure modes rather than a broad claim of frontier model quality. LightningLM 0.1V从稠密种子模型逐步扩展，经5B与9B阶段形成120B稀疏MoE，包含460个路由专家并采用top-12路由，整个流程均在单个八GPU节点完成。118.67B存储参数中仅5.93B处于激活状态，可逆递归结构在反向传播时重建激活值以控制内存。论文明确定位为系统经验报告，其价值在于可复现的扩展方法与失败案例，而非宣称达到前沿模型质量。

AI Models & Inference Systems AI模型与推理系统

Sparsely Gated Single-Neuron Experts 稀疏门控的单神经元专家

S. Schug

arXiv:2606.07414 · 2026-06-05T16:06:47Z

The work shrinks each MoE expert to a single linear neuron and activates only a small fraction of a large expert pool. In isoflop comparisons, replacing Transformer feed-forward layers with these sparsely gated linear neurons improves language-model perplexity across compute budgets. The linear and highly sparse circuits also expose semantically structured expert clusters, though the evidence is still based on relatively small interpretability studies. 该工作将每个MoE专家缩小为单个线性神经元，并只激活大规模专家池中的极少部分。在等FLOP比较中，用这种稀疏门控线性神经元替换Transformer前馈层，可在不同算力预算下改善语言模型困惑度。其线性、高稀疏电路还呈现语义结构化的专家簇，但可解释性证据目前仍主要来自小规模实验。

A CPU-Bound Breakdown of vLLM Cold Starts vLLM冷启动的CPU瓶颈分析

H. S. Kabakibo, A. Trivedi, L. Wang

arXiv:2606.07362 · 2026-06-05T15:07:22Z

This study decomposes vLLM startup into six stages and finds that cold-start latency is predominantly CPU-bound, including in newer configurations using the V1 API and torch.compile. The authors derive an analytical model that predicts startup time for a given hardware configuration and release their profiling data and tools. The result matters for autoscaling and serverless inference, where model execution speed does not compensate for slow process initialization. 该研究将vLLM启动过程拆分为六个阶段，发现冷启动延迟主要受CPU限制，包括采用V1 API与torch.compile的新配置。作者进一步建立可根据硬件配置预测启动时间的解析模型，并开放分析数据与工具。该结果对自动扩缩容和无服务器推理尤为重要，因为模型运行再快也无法抵消缓慢的进程初始化。

Predictive Shortest-Job Scheduling for Serial LLM Backends 面向串行LLM后端的预测式最短作业调度

A. Sundaresan

arXiv:2606.07248 · 2026-06-05T13:19:05Z

Clairvoyant is a sidecar proxy that predicts response length and applies shortest-job-first scheduling to serial OpenAI-compatible backends such as Ollama and llama.cpp. Its XGBoost predictor uses 19 lexical features and adds 0.029 ms per request, targeting head-of-line blocking without the KV-cache memory cost of continuous batching. Ranking accuracy falls from 62-96% in-distribution to 52-66% across distributions, so robustness to workload shift remains the key caveat. Clairvoyant是一个sidecar代理，通过预测响应长度并采用最短作业优先调度，为Ollama和llama.cpp等串行OpenAI兼容后端缓解队首阻塞。其XGBoost预测器使用19个词法特征，每个请求仅增加0.029 ms开销，同时避免连续批处理带来的KV-cache内存成本。排序准确率从分布内的62%至96%下降到跨分布的52%至66%，因此对工作负载变化的鲁棒性仍是主要限制。

Quantum Computing 量子计算

QBugLM Benchmarks LLM Agents on Quantum Software Debugging QBugLM评测LLM智能体的量子软件调试能力

A. B. B. Pham, H. T. Nguyen, M. Usman

arXiv:2606.07314 · 2026-06-05T14:34:46Z

QBugLM automates bug injection, LLM-based detection and repair, and simulation validation for framework-agnostic OpenQASM 3.0 programs. In tests with Claude 4.6 Sonnet and Qwen3 Coder Next, one retry raises Pass@1 from below 25% to above 80%, and simple structured prompts can outperform Chain-of-Thought and ReAct under fixed resources. The large retry effect also shows why quantum-debugging benchmarks must validate repaired programs rather than trust first-pass textual explanations. QBugLM面向框架无关的OpenQASM 3.0程序，自动执行缺陷注入、基于LLM的检测与修复，以及仿真验证。在Claude 4.6 Sonnet与Qwen3 Coder Next测试中，增加一次重试可将Pass@1从低于25%提升至80%以上；在固定资源下，简单结构化提示甚至可优于Chain-of-Thought与ReAct。重试带来的巨大增益也说明，量子调试评测必须验证修复后的程序，而不能只相信首轮文字解释。