Devices & Process
器件与工艺
Xurong Li, Xuan Yang, Z. Qiu, et al.
arXiv:2608.09631 · 2026-08-10
The authors demonstrate a self-starting integrated mode-locked laser using erbium-implanted silicon-nitride gain waveguides, chirped Bragg gratings and a saturable-absorber mirror. It operates from 0.5 to 1.2 GHz with pulses as short as 300 fs and a 27.3 mW mode-locking threshold, targeting a repetition-rate regime that chip-scale sources have struggled to reach.
作者展示了一种自启动全集成锁模激光器，采用掺铒氮化硅增益波导、啁啾布拉格光栅和可饱和吸收镜。该器件工作频率为0.5至1.2 GHz，最短脉冲300 fs，锁模阈值27.3 mW，瞄准片上光源长期难以覆盖的低重复频率区间。
Tianchi Zhang, Yucong Yang, Weihao Yang, et al.
arXiv:2608.09003 · 2026-08-10
This work reports RF-sputtered Ce:YIG films on 4-inch silicon wafers and demonstrates ring-resonator magneto-optical isolators. The films show a 2,318 deg/cm Faraday effect, 80 dB/cm propagation loss and 3.5% thickness uniformity; the result addresses a key scale-up obstacle for nonreciprocal silicon-photonic components.
该工作报告了在4英寸硅晶圆上通过射频溅射制备Ce:YIG薄膜，并演示环形谐振器磁光隔离器。薄膜的法拉第旋转为2318 deg/cm、传播损耗80 dB/cm、厚度均匀性3.5%，直指非互易硅光子器件规模化的一项关键障碍。
Systems & Architecture
系统与架构
Issac Zhu, Hscos Zhang, Ke Jiang, et al.
arXiv:2608.08482 · 2026-08-09
FlashBoot redesigns model bring-up around an exportable contiguous tensor layout, bulk zero-copy host transfers and remote GPU cloning. The paper argues that existing loaders waste fabric bandwidth through fragmented tensors and 10–110 second NCCL setup, making this work relevant to elastic MoE serving where racks must become ready quickly.
FlashBoot围绕可导出的连续张量布局、批量零拷贝主机传输和远程GPU克隆重新设计模型启动流程。论文指出，现有加载器因张量碎片化以及10至110秒的NCCL初始化而浪费互连带宽，因此对需要快速就绪的弹性MoE机架服务具有现实意义。
Yuemeng Xu, Zongxi Liu, Junyu Long, et al.
arXiv:2608.08751 · 2026-08-09
InSituANN keeps base vectors in host memory, performs fine search in place and uses a GPU for compact routing and optional pruning, avoiding PCIe transfers of high-dimensional vectors. On SIFT-1B, it reports a 5.2-minute IVF index build, about 350 times faster than the cited comparison, making retrieval infrastructure less dependent on GPU-resident capacity.
InSituANN将原始向量保留在主机内存中原位执行精细检索，并以GPU处理紧凑路由和可选剪枝，从而避免经PCIe传输高维基础向量。在SIFT-1B上，它报告5.2分钟完成IVF索引构建，约为论文对比方案的350倍，使检索基础设施对GPU常驻容量的依赖更低。
AI Accelerators & Compute-in-Memory
AI加速器与存算一体
Zhongkai Yu, O. Venkatachalam, Zheng Wang, et al.
arXiv:2608.05483 · 2026-08-06
PLoRA proposes a pooled-memory serving design that keeps LoRA adapters and KV cache in the pool, using near-data processing to return reduced results rather than moving all data to the GPU. It addresses deployments with thousands of adapters, where CPU-DRAM staging over PCIe consumes both capacity and kernel-launch overhead; its benefit depends on the availability of memory-semantic fabrics and NDP support.
PLoRA提出一种池化内存服务设计，将LoRA适配器和KV缓存留在内存池中，通过近数据处理只返回归约结果，而非把全部数据搬到GPU。它针对拥有数千个适配器的部署场景，解决CPU DRAM经PCIe调入所造成的容量和内核启动开销；其收益仍取决于具备内存语义的互连和NDP支持。
Wentao Zhong, Yida Dong, Wenlu Zhong, et al.
APL Photonics · 2026-08-01
A GST superlattice on a silicon microring is used to demonstrate five reconfigurable nonlinear activation functions, including ReLU and softplus. The device reaches a 1.76 mW activation threshold and up to 1 MHz operation; a system-level MNIST experiment improves reported accuracy from 95.35% to 99.24%, though that benchmark does not establish an end-to-end optical-compute advantage.
研究在硅微环上集成GST超晶格，演示了包括ReLU和softplus在内的五种可重构非线性激活函数。器件激活阈值为1.76 mW、最高工作速度1 MHz；系统级MNIST实验报告准确率从95.35%升至99.24%，但这一基准尚不足以证明端到端光计算优势。
AI Systems & Inference
AI系统与推理
Can Xiao, Sukmin Cho, Junbong We, et al.
arXiv:2608.08097 · 2026-08-08
OasisKV decouples full KV-cache storage from HBM during LLM decoding, retaining only predicted-relevant entries in HBM and prefetching blocks from higher-capacity tiers. Built on vLLM, it uses speculative-decoding lookahead to predict future important tokens, offering a systems path to larger contexts without assuming that all KV state fits on the accelerator.
OasisKV在LLM解码期间将完整KV缓存存储与HBM解耦，只将预测为相关的条目保留在HBM中，并从更大容量的层级预取数据块。该系统基于vLLM，利用推测解码的前瞻信息预测未来重要token，为更长上下文提供了一条不要求所有KV状态都装入加速器的系统路径。
Zhiqiang Xie, Zhangheng Huang, Ting-Jun Huang, et al.
arXiv:2608.07009 · 2026-08-07
HiSparse keeps full KV history in host memory while bounding each request's GPU footprint with a fixed-size cache, preserving exact model outputs for sparse-attention methods. Its fused CUDA path combines hit detection, replacement and host-device fetches inside the decode graph, and the authors evaluate it on H200, B200 and GH200 systems; the approach is most applicable when sparse selection remains effective.
HiSparse将完整KV历史保留在主机内存中，同时以固定大小的GPU缓存限制每个请求的显存占用，并保持稀疏注意力方法的模型输出精确一致。其融合CUDA路径在解码图内完成命中检测、替换和主机至设备取数，作者在H200、B200和GH200系统上评估；该方法最适合稀疏选择仍有效的场景。
Vincent-Daniel Yun, Woosang Lim, Minsoo Cheong, et al.
arXiv:2608.02691 · 2026-08-03
OptR learns per-head orthogonal corrections for INT2 KV-cache quantization by minimizing error after attention and the output projection, rather than optimizing a pre-output proxy. Across three models and five reasoning and coding benchmarks, it improves on QuaRot and OSCAR while retaining a paged KV-cache format with negligible stated overhead, making the method relevant where two-bit cache compression is necessary.
OptR通过最小化注意力和输出投影之后的误差，为INT2 KV缓存量化学习逐头正交校正，而不是优化输出前的代理误差。它在三种模型、五项推理与编程基准上优于QuaRot和OSCAR，同时保留分页KV缓存格式且声称开销可忽略，适用于必须采用2比特缓存压缩的场景。