AI Accelerators & Compute-in-Memory
AI加速器与存算一体
M. Fadel Argerich, J. Fuerst, M. Patino-Martinez
arXiv:2607.02391 · 2026-07-02T16:25:34Z
WattGPU predicts mean GPU power draw and inter-token latency for LLM inference using only public model metadata and GPU specifications. On 42 open-source LLMs and 8 NVIDIA server GPUs, the paper reports median absolute percentage error at or below 3.4% for offline power and 13.5% for server-mode power on unseen GPUs, with server-mode latency error at or below 8.5%. If it generalizes, this gives operators a planning tool before they spend time profiling every model-and-GPU pair.
WattGPU只使用公开模型元数据和GPU规格，预测LLM推理的平均GPU功耗和token间延迟。论文在42个开源LLM和8款NVIDIA服务器GPU上报告，面对未见过的GPU，离线功耗中位绝对百分比误差不超过3.4%，服务器模式功耗不超过13.5%，服务器模式延迟误差不超过8.5%。如果泛化成立，它能让运营方在逐一实测模型与GPU组合之前先完成容量和能耗规划。
J. Shen, Y. Wu, C. Chen
arXiv:2607.02283 · 2026-07-02T15:02:28Z
Shen, Wu, and Chen argue that a single dendritic compartment can implement the online learning dynamics needed for in-context learning in spiking neural networks. Their DendriCL architecture maps apical recurrence to a leaky Widrow-Hoff LMS update, reducing what would normally require deeper transformer-like machinery. The result is promising for neuromorphic architecture thinking, but the evidence is algorithmic; it is not yet a measured silicon macro.
Shen、Wu和Chen认为，单个树突隔室即可实现脉冲神经网络中上下文学习所需的在线学习动力学。其DendriCL架构把顶端递归映射为带泄漏的Widrow-Hoff LMS更新，从而压缩通常需要更深Transformer式机制才能完成的适应过程。这个结果对神经形态架构有启发，但证据仍是算法层面的，并非已测硅宏单元。
AI Research for Systems
面向系统的AI研究
D. Lee, J. Chavan, D. Nguyen, et al.
arXiv:2607.02461 · 2026-07-02T17:27:34Z
OrbitQuant removes calibration-data dependence from post-training quantization for diffusion transformers by rotating activations into a normalized basis with a randomized permuted block-Hadamard transform. The same Lloyd-Max codebook can then serve many timesteps, prompts, layers, and even image-to-video transfers, according to the paper. The hardware relevance is direct: lower-bit DiT inference is only useful if quantization avoids per-checkpoint calibration overhead and preserves quality across long sampling chains.
OrbitQuant通过随机置换块Hadamard变换，把扩散Transformer的激活旋转到归一化基空间，从而去掉后训练量化对校准数据的依赖。论文称，同一个Lloyd-Max码本可服务多个时间步、提示、层，甚至从图像模型迁移到视频模型。其硬件相关性很直接：低比特DiT推理只有在避免逐检查点校准开销并保持长采样链质量时才真正有用。
X. Wang, X. Yang, W. Shen
arXiv:2607.02484 · 2026-07-02T17:50:57Z
EADP targets the failure mode where visual token pruning drops fine-grained cues under dense multimodal instructions. It estimates textual noise with entropy, then uses submodular selection with a spatial prior instead of naive top-k token retention. For accelerator workloads, the paper is relevant because visual-token count is a first-order driver of memory traffic and attention compute in VLM inference.
EADP针对视觉token剪枝在密集多模态指令下丢失细粒度线索的问题。它先用熵估计文本噪声，再以带空间先验的子模选择替代简单Top-K保留。对加速器负载而言，这篇论文相关性很高，因为视觉token数量直接决定VLM推理中的内存流量和attention计算量。
W. Cui
arXiv:2607.02303 · 2026-07-02T15:19:49Z
HOLA augments linear attention with a bounded exact KV cache so facts that do not compress well into the recurrent state can still be recalled. At 340M parameters trained on 15B SlimPajama tokens, the paper reports Wikitext perplexity improving from 27.32 to 22.92, beating a full-attention Transformer++ baseline at 26.88. The approach is hardware-relevant because it keeps the linear-attention memory model while adding a controlled cache path for high-value associations.
HOLA在线性attention之外加入有界精确KV cache，让无法良好压缩进递归状态的事实仍可被召回。在使用15B SlimPajama token训练的340M参数模型上，论文报告Wikitext困惑度从27.32降至22.92，优于26.88的全attention Transformer++基线。该方法对硬件有意义，因为它保留线性attention的内存模型，同时为高价值关联增加受控cache路径。
M. Li, R. Mittal, S. Rana, et al.
arXiv:2607.02338 · 2026-07-02T15:44:43Z
Li et al. propose a certify-then-rectify framework for HNSW nearest-neighbor search, adding a statistical certifier that escalates uncertain queries to a rigorous recovery path. They reinterpret the HNSW graph as a geometric spanner and use extreme-value estimates to bound where true nearest neighbors can lie. The systems angle is clear: vector retrieval for RAG and memory systems needs predictable worst-case behavior, not only high average throughput.
Li等人提出面向HNSW近邻搜索的“先认证、再修正”框架，加入统计认证器，在查询结果不可靠时升级到严格恢复路径。他们把HNSW图重新解释为几何spanner，并用极值估计限定真实近邻可能出现的位置。系统意义很明确：RAG和记忆系统中的向量检索不仅需要高平均吞吐，还需要可预测的最坏情况行为。