Inference Efficiency Moves Into the Memory Path 推理效率优化深入存储路径

This reading queue concentrates on the cost of moving state: KV-cache reuse, weight offloading, visual-token budgets, on-chip links, and encrypted-inference traffic. A 2D CFET thermal study and a field-ready NV-diamond sensor broaden the hardware view. 本期论文集中关注状态搬运成本，包括KV-cache复用、权重卸载、视觉token预算、片上链路与加密推理流量。2D CFET热设计研究与可现场部署的NV金刚石传感器则扩展了硬件视野。

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

Devices & Process 器件与工艺

AI-Guided Electrothermal Design for 2D CFETs AI辅助的2D CFET电热协同设计

M.-H. Kim, K. Sharma, S. Zhang, et al.

arXiv:2609.17123 · 2026-09-15

An AI-agent workflow proposes a redistributed source interconnect and a substrate-directed heat-removal path for a modeled 12-nm 2D CFET inverter. Together they reduce peak temperature rise by 1.67 K at fixed metal volume and 20 μW; a resistance sensitivity study finds about 0.6 K of cooling with a 2% nFET on-current penalty. The result is an electrothermal-model study rather than measured silicon, so the main contribution is the constrained design-and-verification workflow, not proof of manufacturability. 该研究让AI智能体为一个12 nm 2D CFET反相器模型提出重新分配的源极互连，以及面向衬底的散热路径。在金属体积与20 μW功耗不变时，组合设计把峰值温升降低1.67 K；电阻敏感性分析还显示，可用2%的nFET导通电流损失换取约0.6 K降温。结果来自电热模型而非实测硅片，因此主要贡献是受约束的设计与验证流程，而非可制造性的证明。

Circuits & Architecture 电路与架构

Budgeted Express-Mesh Co-Designs Links and Adaptive Routing Budgeted Express-Mesh协同设计链路与自适应路由

L. Cao, J. Ma

arXiv:2609.17057 · 2026-09-15

Budgeted Express-Mesh adds a limited number of traffic-aware express links under a fixed wiring budget, then jointly selects their placement and deadlock-free adaptive routes. Simulations across four synthetic workloads improve high-load throughput over a conventional mesh and random link placement, with gains persisting under delayed quantized congestion signals and longer link latency. The evidence is simulation-only and workload-sensitive, so area, timing, and physical-design costs remain open. Budgeted Express-Mesh在固定布线预算下加入少量流量感知的快速链路，并协同优化链路位置与无死锁自适应路由。对4种合成负载的仿真显示，其高负载吞吐量优于传统Mesh和随机链路布局；在量化拥塞信号延迟与链路时延增加时，优势仍然存在。现有证据仅来自仿真且依赖负载，面积、时序与物理设计成本仍待验证。

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

OptiPrime Co-Designs Protocols and Hardware for Private Inference OptiPrime协同优化隐私推理协议与硬件

J. Yu, Y. Yu, S. Chen, et al.

arXiv:2609.16898 · 2026-09-15

OptiPrime attacks a system bottleneck that remains after homomorphic-encryption accelerators speed up individual operations: repeated wireless transfer of input and output ciphertexts in HE-MPC inference. Its convolution protocol reduces output ciphertext traffic, while the hardware design is reorganized for the resulting weight-plaintext and intermediate-ciphertext memory pressure. The preprint usefully shifts evaluation from isolated crypto kernels to end-to-end inference, but broader network and model coverage will be needed to establish generality. OptiPrime针对同态加密加速器提升单项运算后仍然存在的系统瓶颈：HE-MPC推理中输入与输出密文的反复无线传输。其卷积协议减少输出密文流量，同时重新组织硬件，以应对由权重明文和中间密文带来的内存访问压力。该预印本把评估重点从孤立的密码学内核转向端到端推理，但仍需覆盖更多网络环境与模型才能证明普适性。

AI Research & Inference Systems AI研究与推理系统

Reusing Shared-Prefix KV Caches Across Standard LoRA Adapters 在标准LoRA适配器之间复用共享前缀KV Cache

D. Rajput

arXiv:2609.17109 · 2026-09-15

The paper tests whether one backbone prefill KV cache can be reused across already-trained LoRA specialists instead of recomputing a shared context for each adapter. On Qwen3-1.7B with HotpotQA and GSM8K adapters, full-prefix reuse minimizes prefill cost but changes GSM8K exact match by -4.6 points at a 160-token budget, -3.0 at 320 tokens, and -0.8 under a second seed. Partial recomputation shows no demonstrated advantage, and the limited model-task sweep does not establish quality equivalence or a universal handoff boundary. 论文测试能否在多个已训练LoRA专家之间复用一次骨干模型的前缀KV Cache，从而避免为每个适配器重复计算相同上下文。在搭载HotpotQA与GSM8K适配器的Qwen3-1.7B上，全前缀复用把预填充成本降到最低，但GSM8K精确匹配率在160-token预算下变化-4.6个百分点、320-token下为-3.0，第二个随机种子下为-0.8。局部重算未显示优势，而有限的模型与任务范围也不足以证明质量等价或给出通用切换边界。

FlexEE Makes Early Exit Compatible with KV Caches and Weight Offloading FlexEE让提前退出兼容KV Cache与权重卸载

Q. Xie, Z. Li, Y. Kang

arXiv:2609.17008 · 2026-09-15

FlexEE combines layer-wise exit supervision, self-speculative decoding over a local Top-K vocabulary, and dynamic hidden-state management so early exits remain KV-cache correct. It reports up to 1.27× and 3.16× end-to-end speedups on Llama2-7B at 0% and 50% weight offloading, and 1.25× and 2.83× on Llama3-8B, with minimal stated accuracy loss. The largest gains arrive in offloaded settings, so performance on fully resident production accelerators needs separate validation. FlexEE结合逐层退出监督、局部Top-K词表上的自推测解码，以及动态隐藏状态管理，使提前退出仍能保持KV Cache正确。论文报告称，Llama2-7B在0%与50%权重卸载时端到端最高加速分别为1.27×和3.16×，Llama3-8B则为1.25×和2.83×，且准确率损失很小。最大收益出现在权重卸载场景，因此仍需单独验证模型完全驻留于生产加速器时的表现。

StackTok Adapts Visual-Token Selection to Query and Budget StackTok按查询与预算自适应筛选视觉token

Z. Wang, L. Zhang, L. Wang, et al.

arXiv:2609.16841 · 2026-09-15

StackTok is a training-free visual-token selector that adjusts the balance between query relevance and image coverage according to the token budget and query-vision affinity. For high-resolution images, it also allocates a shared budget across crops using marginal gains from locally nominated tokens. Evaluation spans five VLMs and ten image-understanding benchmarks, but the preprint still needs deployment-level latency and memory measurements to show how token savings translate into hardware gains. StackTok是一种无需训练的视觉token筛选器，会根据token预算及查询与图像的相关性，自适应平衡查询相关度和画面覆盖度。对于高分辨率图像，它还依据各裁剪区域候选token的边际收益分配统一预算。论文覆盖5个VLM与10项图像理解基准，但仍需部署级延迟与内存测量，才能说明token节省能转化为多大硬件收益。

Quantum Devices & Systems 量子器件与系统

Portable NV-Diamond Magnetometer Reaches Shot-Noise-Limited Operation 便携式NV金刚石磁力计达到散粒噪声极限

A. K. P, V. Rane, S. Atakar, et al.

arXiv:2609.16901 · 2026-09-15

A portable ensemble NV-diamond magnetometer combines dynamic differential readout, in-situ lineshape tracking, and a magnetic eigenvector transform to suppress laser noise and thermal drift. In unshielded conditions it reaches an off-resonance noise factor of 1.0 ± 0.1 and cuts long-term drift tenfold. That moves vector quantum sensing closer to field use, although durability and calibration stability outside the reported setup remain to be demonstrated. 一款便携式系综NV金刚石磁力计结合动态差分读出、原位线形跟踪与磁本征向量变换，用于抑制激光噪声和温度漂移。在无磁屏蔽环境下，系统实现1.0 ± 0.1的离共振噪声因子，并把长期漂移降低10倍。该结果让矢量量子传感更接近现场应用，但在所报告装置之外的耐久性与校准稳定性仍待证明。