Accelerator Systems & Interconnects
加速器系统与互连
Lei Tian, Wenhao Gu
arXiv:2611.21288 · 2026-11-16T15:55:18Z
GNoF moves NVMe-oF/RDMA command construction, submission and completion handling onto the GPU, retaining CPU involvement for connection setup and administrative tasks. It batches doorbell updates and completion polling to reduce the SM time consumed while waiting through 10–100 microsecond fabric round trips. This targets a practical cost of GPU-initiated remote storage: freeing the CPU is insufficient if storage waits occupy GPU compute resources.
GNoF将NVMe-oF/RDMA命令构造、提交和完成处理移到GPU上,CPU仍负责连接初始化和管理任务。系统批量处理门铃更新与完成队列轮询,以减少等待10至100微秒网络往返时占用的SM资源。这项工作针对GPU主动访问远程存储的一项实际代价:若存储等待持续占用GPU计算资源,仅释放CPU还不够。
Richard Henwood, John Lockman, Luke Wilson
arXiv:2611.21442 · 2026-11-16T17:30:44Z
The Bipartible architecture places clones of a global HTTP service on HPC compute hosts so storage requests can use the InfiniBand network without modifying workload code or network configuration. It addresses the mismatch between cloud-native HTTP applications and clusters whose high-performance storage fabric is separate from TCP networking. The approach assumes the persistent storage is reachable over InfiniBand, so its applicability depends on the cluster layout.
Bipartible架构在HPC计算主机上部署全局HTTP服务的副本,让存储请求利用InfiniBand网络,而无需修改工作负载代码或网络配置。它针对云原生HTTP应用与HPC集群之间的适配问题,后者的高性能存储网络通常独立于TCP网络。该方法要求持久化存储可通过InfiniBand访问,因此适用范围取决于集群布局。
Fan Yang, Xiaolong Wu, Qi Cao, et al.
arXiv:2611.20824 · 2026-11-16T10:06:22Z
LPAS synthesizes a distributed DNN architecture from the physical network topology under communication, compute, storage, synchronization and latency constraints. Across the evaluated networks, the authors report gaps below 0.6% to an LP-derived upper bound and lower runtime than direct mixed-integer optimization. This gives designers a way to fit neural computation to available infrastructure, although proximity to an optimization bound does not establish model accuracy.
LPAS在通信、计算、存储、同步和延迟约束下,依据物理网络拓扑合成分布式DNN架构。作者报告,在所测试的网络上,求解结果与线性规划推导的上界相差不到0.6%,运行时间也低于直接采用混合整数优化的方法。这为按现有基础设施安排神经网络计算提供了路径,但接近优化上界并不等于已经证明模型精度。
AI Training, Architectures & Inference
AI训练、架构与推理
Steve Rhyner, Yubo Gao, Stefan de Lasa, et al.
arXiv:2611.21584 · 2026-11-16T18:54:07Z
CertiSum certifies when native FP32 accumulation of MXFP4 products is exact and recomputes uncertified results, while keeping quantization blocks and partial-sum merging invariant across layouts. In 200-step Qwen3-8B-Base experiments, training states matched bit for bit across AMD MI350X and Nvidia GB200 and across parallelism changes on MI350X. Across three 8B–14B models, the reported training-step latency overhead was 11.8–60.5% versus non-exact MXFP4, making reproducibility an explicit performance trade-off.
CertiSum验证原生FP32累加MXFP4乘积何时能够保持精确,对未通过验证的结果重新精确计算,并使量化块和部分和合并不随并行布局改变。在Qwen3-8B-Base的200步实验中,AMD MI350X与英伟达GB200上的训练状态逐比特一致,在MI350X上改变并行方式也保持一致。对3个80亿至140亿参数模型,作者报告相较非精确MXFP4,单步训练延迟增加11.8%至60.5%,说明可复现性需要付出明确的性能代价。
Yifan Sun, Jiantong Jiang, Zhexiang Zhang, et al.
arXiv:2611.21165 · 2026-11-16T14:29:27Z
Across four multimodal models with a fixed three-GPU budget, separating visual encoding, prefill and decode reduced GPU energy per successful request by 13.3–37.0% at nearly unchanged output-token throughput, but increased mean time to first token. Across the evaluated workloads, uniform downclocking to 80% reduced GPU energy by 11.7–18.6% with modest changes in mean latency. The measurements show why stage-level power savings must be evaluated with pipeline queueing and request latency, rather than power alone.
在固定使用3颗GPU、测试4个多模态模型的条件下,将视觉编码、预填充和解码拆分,使每个成功请求的GPU能耗降低13.3%至37.0%,输出词元吞吐量基本不变,但平均首词元延迟上升。在所测试的负载中,统一降频至80%可使GPU能耗降低11.7%至18.6%,平均延迟变化较小。这些测量说明,评估单个阶段的节电收益时,需要同时考虑流水线排队与请求延迟,不能只看功率。
Bryan A. Plummer, Shan Lu, Keval Doshi, et al.
arXiv:2611.21459 · 2026-11-16T17:43:21Z
Balanced Twin Weight Factorization reuses a single shared-basis forward pass when factorized layers receive the same input. Its balancing mechanism aims to prevent larger layers from dominating the shared representation when the paired layers have different sizes. The method targets the gap between reducing FLOPs and reducing actual latency, where multiple small matrix operations can otherwise erase compression benefits.
平衡孪生权重分解在多个分解层接收相同输入时,只执行一次共享基矩阵的前向计算并复用结果。当成对的层尺寸不同时,其平衡机制旨在避免较大的层主导共享表示。该方法针对减少FLOPs却未必降低实际延迟的问题,因为用多个小矩阵运算替代一次大矩阵运算,可能抵消压缩带来的收益。
NVIDIA, Adi Margolin, Aditya Malte, et al.
arXiv:2611.20771 · 2026-11-16T09:21:45Z
Nemotron 3.5 Bolt E4B 4GB further post-trains Gemma-4-E4B-it for conversational and tool-using agents and includes a DFlash draft model. The authors report 2.74× per-user decode speed with DFlash versus decoding without speculation at concurrency one, and 1.21× the generation throughput of Gemma-4-E4B-it with MTP at batch size one. These are distinct baselines and single-user or single-batch conditions, so the figures should not be read as general serving-throughput gains.
Nemotron 3.5 Bolt E4B 4GB在Gemma-4-E4B-it基础上进一步后训练,面向对话和工具调用智能体,并配套DFlash草稿模型。作者报告,在并发数为1时,使用DFlash的单用户解码速度为不采用投机解码时的2.74倍;在批大小为1时,生成吞吐量为采用MTP的Gemma-4-E4B-it的1.21倍。两项结果对应不同基线,且均为单用户或单批次条件,不能直接视为普遍适用的服务吞吐增益。
Shivam Duggal, Zongze Wu, William T. Freeman, et al.
arXiv:2611.21608 · 2026-11-16T18:59:38Z
Latent Lattice Language Models add per-layer recurrent latent streams that persist across tokens and condition attention, KV-cache writes, MLP computation and decoding. After training on 20 billion FineWeb-Edu tokens, a 406-million-parameter model reduced validation perplexity from 15.41 to 14.40 versus a parameter-matched Transformer. The architecture changes the state and computation that inference hardware must support, but the reported quality gain does not establish a latency or energy improvement.
潜变量格点语言模型在各层增加跨词元持续存在的循环潜变量流,用于调节注意力、KV缓存写入、MLP计算和解码。在200亿个FineWeb-Edu词元上训练后,一个4.06亿参数模型的验证困惑度从参数量相当的Transformer的15.41降至14.40。这一架构改变了推理硬件需要支持的状态和计算方式,但所报告的质量提升并不等于已证明延迟或能耗改善。