AI Accelerators & Compute-in-Memory
AI加速器与存算一体
W.-H. Huang, K. Lee, M.-Y. Lee, et al.
arXiv:2609.01084 · 2026-09-01T11:18:55Z
This work co-designs wide-I/O LPDDR, precision-tagged KV reads, feed-forward reuse, and a mixed-precision systolic array for block-diffusion LLMs at batch one. On modeled Jetson-class platforms with 1.5B and 7B models, the authors report mean energy reductions of 3.79x and 3.96x and latency speedups of 2.88x and 4.44x, with benchmark losses below one percentage point. The gains are simulation-based rather than measured silicon, and several techniques depend on the execution properties of native block diffusion.
该工作针对batch-one的block-diffusion LLM，协同设计宽I/O LPDDR、带精度标签的KV读取、前馈网络复用和混合精度脉动阵列。在建模的Jetson级平台上，1.5B与7B模型分别实现平均3.79倍和3.96倍能耗降低，以及2.88倍和4.44倍延迟加速，基准分数下降小于1个百分点。结果来自仿真而非实测硅片，且多项技术依赖原生block diffusion的执行特性。
Y. Wang, Y. J. Chong, X. Fong
arXiv:2609.00857 · 2026-09-01T07:55:45Z
LEAP maps static weights to in-memory-compute processing elements, dynamic intermediates to near-memory compute, and partial-result reduction to in-network compute. A companion partitioning and scheduling layer changes the mapping for the distinct arithmetic intensity of prefill and decode. The framework attacks data movement across a scale-up fabric, but the preprint describes a co-design proposal rather than a measured chip.
LEAP把静态权重映射到存内计算处理单元，把动态中间数据交给近存计算，并用网内计算完成部分结果归约。配套的分区与调度层针对prefill和decode不同的算术强度调整映射。该框架瞄准scale-up互连中的数据移动，但预印本描述的是协同设计方案，并非实测芯片。
Hardware-Relevant AI Research
硬件相关AI研究
W. Wen, Y. Liu, Y. Liu, et al.
arXiv:2609.01068 · 2026-09-01T11:00:39Z
OUTLETS reuses hidden representations already produced by a speculative-decoding draft model to predict generation length with only a lightweight regression head. The authors report lower mean absolute error than evaluated proxy and internal-state methods, then use the predictions to improve scheduling under saturated disaggregated serving. The systems benefit depends on already running a compatible speculative backbone and on workload length distributions matching the evaluation.
OUTLETS复用推测解码draft model已经生成的隐藏表示，仅增加轻量回归头来预测生成长度。作者报告其平均绝对误差低于所评估的代理模型和内部状态方法，并利用预测结果改善饱和状态下的解耦式服务调度。系统收益取决于是否已运行兼容的推测解码骨干，也取决于实际工作负载的长度分布是否与评估相符。
Z. Gan, F. Liu, C. Guan, et al.
arXiv:2609.01024 · 2026-09-01T10:21:10Z
PCoMoE decomposes MoE experts into reusable sub-expert paths, then prunes compatible path combinations and executes them with a hardware-oriented runtime. The paper reports up to 1.31x end-to-end inference speedup while also reporting a 10% accuracy improvement over its comparison point. Because both gains come from a preprint evaluation, the accuracy baseline and runtime overhead deserve close inspection before generalizing the result.
PCoMoE把MoE专家拆分为可复用的子专家路径，再剪除低价值的兼容路径组合，并用面向硬件的运行时执行。论文报告端到端推理最高加速1.31倍，同时相对其比较基线提高10%的准确率。由于两项增益都来自预印本评估，在推广结论前需要仔细核对准确率基线和运行时开销。
N. Xiros, D. Damianos, M.-E. Zoumpoulidi, et al.
arXiv:2609.01100 · 2026-09-01T11:43:41Z
CoRM routes each token by contrasting it with an exponential-moving-average reference state, aiming to remove features shared across tokens and expose a more separable routing subspace. Across nine zero-shot reasoning benchmarks, the authors report gains of 0.67 to 1.69 points for Top-1 routing and 1.38 to 1.77 points for Top-2, at costs of 2.9% more parameters and 2.6% more FLOPs per token. The results improve specialization, but the added compute and benchmark-only evidence leave serving efficiency unproven.
CoRM把每个token与指数移动平均参考状态进行对比式路由，以滤除token之间共享的特征，暴露出更易分离的路由子空间。在9项zero-shot推理基准上，作者报告Top-1路由提高0.67至1.69分、Top-2提高1.38至1.77分，代价是参数量增加2.9%、每token FLOPs增加2.6%。结果显示专家分工有所改善，但额外计算成本和仅基于基准测试的证据尚不足以证明实际服务效率。
S. Li, Z.-Y. Hu, S. Huang, et al.
arXiv:2609.01004 · 2026-09-01T09:52:10Z
SinkPruner identifies high-norm visual tokens as redundant attention sinks, removes them with a visual sanitizer, and then retains query-relevant tokens with a text-guided stage. The training-free method is evaluated across 12 image-language and four video-language benchmarks, making it directly relevant to multimodal inference cost. The abstract does not expose the full accuracy-throughput trade-off, so deployment value depends on how aggressively tokens can be pruned without task-specific regressions.
SinkPruner把高范数视觉token识别为冗余的attention sink，先通过视觉清洗器移除，再用文本引导阶段保留与查询相关的token。该方法无需训练，并在12项图文和4项视频语言基准上评估，直接对应多模态推理成本。摘要没有给出完整的准确率—吞吐权衡，因此部署价值取决于在不引发任务特定退化的情况下能剪除多少token。
J. Liu, S. Ranjan, P. Mishra, et al.
arXiv:2609.00791 · 2026-09-01T06:38:17Z
Instella-MoE is a fully open model with 16 billion total parameters and 2.8 billion active parameters per token, trained from scratch on AMD MI300X and MI325X GPUs. Its design combines Gated Multi-head Latent Attention with FarSkip-Collective connectivity, and the report claims a 76.7 average across its pre-training benchmark suite. The technical report is useful as a rare open AMD training-system case study, but aggregate benchmark averages do not reveal cluster efficiency or production-serving behavior.
Instella-MoE是一款完全开放的MoE模型，总参数160亿、每token激活28亿参数，并从零开始在AMD MI300X和MI325X GPU上训练。其设计结合Gated Multi-head Latent Attention与FarSkip-Collective连接，报告在预训练基准套件上的平均分为76.7。该技术报告是少见的开放AMD训练系统案例，但综合基准均分无法说明集群效率或生产服务表现。