Tagged articles
large-scale training
26 articles · Page 1 of 1
H3C's Wang Feng details a full-stack super-node architecture addressing AI infrastructure challenges with 128 GPUs per cabinet via full liquid cooling, 96% network utilization through optical-electrical parallel links, 200GB/s storage bandwidth, 30-day stable training at 100k nodes, three-layer security, and AI-driven operations cutting deployment time 60% and fault diagnosis 90%.
AI infrastructureAIOpsH3C
0 likes · 18 min read

The article examines the PhoneWorld project, which reconstructs realistic Android app environments from user interaction traces to create scalable, resettable, and verifiable mock apps, enabling large‑scale training and evaluation of Mobile Agents with demonstrated performance gains across multiple benchmarks.
Mobile AgentPhoneWorldapp generation
0 likes · 12 min read

The D2AR generative compression framework leverages historical Earth‑observation priors and a dual‑decoupled asymmetric design to achieve exascale training at 2.16 EFLOP/s on a domestic Armv9 CPU supercomputer, scaling to 20,480 nodes and delivering up to 10,000× data reduction while preserving scientific utility.
Armv9 supercomputerdata reductionexascale AI
0 likes · 8 min read

GIPO (Gaussian Importance Sampling Policy Optimization) replaces PPO’s hard clipping with a smooth Gaussian‑weighted trust region, achieving log‑space symmetry and bias‑variance balance that mitigates policy lag and utilization collapse, and demonstrates superior stability and sample efficiency on GridWorld, LIBERO, MetaWorld, and 7‑billion‑parameter VLA experiments.
Bias-Variance TradeoffGIPOPolicy Optimization
0 likes · 17 min read

Agent-World introduces a universal training arena that automatically mines real‑world data from the internet to build over 1,900 diverse environments and 19,800 tools, then generates long‑horizon tasks through graph‑based and programmatic synthesis, creating a self‑evolving loop where agents are evaluated, diagnosed, and the environment is refined, achieving state‑of‑the‑art results on 23 benchmarks.
AI agentsAgent-Worldbenchmark evaluation
0 likes · 14 min read

This article details JD Retail’s 9N‑LLM unified training engine, explaining the background of generative recommendation, the challenges of massive sparse and dense parameters, and the multi‑framework, multi‑hardware solutions—including efficient sample processing, large‑scale sparse embedding, dense scaling, UniAttention acceleration, and reinforcement‑learning integration—that enable industrial‑scale deployment.
AI infrastructureSparse Embeddinggenerative recommendation
0 likes · 26 min read


ArchitectJan 1, 2026 · Artificial Intelligence
DeepSeek’s new paper introduces mHC, a manifold‑constrained version of Hyper‑Connections that stabilizes gradient flow, adds only 6.7% training overhead, and enables reliable training of 27‑billion‑parameter models while improving benchmark performance by about 2%.
AI ArchitectureHyper-ConnectionsManifold-Constrained
0 likes · 7 min read

The article explains how Google’s TPU v7 supercomputer uses a simple yet powerful networking scheme—1.5 optical modules per TPU for intra‑rack communication and an additional 2.6 modules per TPU for inter‑rack high‑speed links—enabling massive AI model training with balanced cost and performance.
AI supercomputingGoogleTPU
0 likes · 13 min read

In a deep interview, Baidu AI Computing chief scientist Wang Yanpeng and host Koji trace China's internet infrastructure from the early big‑data era through cloud computing to today's AI boom, highlighting the pivotal role of compute power, GPU acceleration, data scaling, and Baidu's Baige platform in shaping the AI arms race.
AI infrastructureBaidu BaigeGPU computing
0 likes · 26 min read

This article examines how AI large‑model training demands massive GPU clusters and low‑latency, high‑throughput networks, compares Clos/Fat‑Tree, Spine‑Leaf, Dragonfly, Group‑wise Dragonfly+ and Torus topologies, and discusses design choices for scaling to hundreds of thousands of GPUs while noting related data‑center resources.
AIdata centerlarge-scale training
0 likes · 8 min read

Tencent's LFM4Ads foundation model introduces a full-representation, multi-granular knowledge transfer framework that moves user, item, and cross representations to downstream tasks, dramatically improving ad recommendation metrics across dozens of business scenarios.
Representation Learningad recommendationfoundation model
0 likes · 10 min read

Meta's DINOv3 is a 70‑billion‑parameter self‑supervised visual foundation model trained on 17 billion Instagram images without any labels, introducing dense feature extraction, Gram‑Anchoring to prevent feature collapse, high‑resolution adaptation, and multi‑student distillation that together enable out‑of‑the‑box performance on segmentation, depth estimation, 3D matching, and tracking while surpassing prior models such as DINOv2, CLIP, and SAM.
Computer VisionDINOv3Gram Anchoring
0 likes · 8 min read

Meituan’s recommendation team introduced the MTGR framework, aligning traditional DLRM features with a unified HSTU‑based Transformer to explore scaling laws, delivering a 65‑fold FLOPs boost, 12% lower inference cost, and record gains in online CTR and order volume across its food‑delivery platform.
Inference OptimizationMTGRScaling Law
0 likes · 26 min read

The article explains the core concepts and trade‑offs of five parallelism techniques—data, tensor, context, pipeline, and expert parallelism—plus the ZeRO optimizer, showing when each method is appropriate for training ultra‑large PyTorch models and providing concrete code snippets and performance considerations.
Context ParallelismExpert ParallelismPipeline Parallelism
0 likes · 21 min read

The article traces the evolution of AI training stability from early manual operations on small GPU clusters to sophisticated, fault‑tolerant infrastructures for thousand‑card and ten‑thousand‑card models, detailing Baidu Baige’s metrics, monitoring, eBPF‑based diagnostics, and checkpoint strategies that reduce invalid training time and accelerate fault recovery.
checkpointingdistributed systemseBPF
0 likes · 22 min read


AIWalkerFeb 12, 2025 · Artificial Intelligence
The paper presents Goku, a rectified‑flow transformer that jointly generates high‑quality images and videos at commercial scale, detailing its novel architecture, massive high‑quality data pipeline, efficient large‑scale training tricks, and state‑of‑the‑art results on GenEval, DPG‑Bench, VBench and UCF‑101.
Image Generationflow-based modelslarge-scale training
0 likes · 29 min read

This article explains how Tencent’s AI team leverages the PyTorch‑based TorchRec library and a custom dynamic embedding solution to train billion‑scale recommendation models efficiently, detailing the benefits of TorchRec, GPU embedding, optimized kernels, embedding partition strategies, experimental results, and practical deployment guidance.
GPU EmbeddingPyTorchTorchRec
0 likes · 15 min read

This article introduces the challenges of scaling deep‑learning model training, explains the design and components of the open‑source Easy Parallel Library (EPL) that unifies data, pipeline, and operator‑split parallelism, and demonstrates its best‑practice results on large‑scale classification, BERT‑large, and massive multimodal models.
EPLParallelismZeRO
0 likes · 15 min read

This article presents Tencent's latest advancements in large‑scale model training, detailing the AngelPTM framework and its ZeRO‑Cache optimization techniques that reduce memory and storage costs, improve hardware utilization, and achieve high‑performance training for trillion‑parameter AI models across various applications.
AI modelsAngelPTMZeRO-Cache
0 likes · 14 min read

PGLBox, Baidu’s GPU‑based large‑scale graph training framework, delivers up to 27× speedup over CPU clusters by fully GPU‑accelerating storage, sampling, and training, supporting billions of nodes, advanced GNN algorithms, multi‑level storage, and seamless integration of massive pretrained models.
GPUGraph Neural NetworksPGLBox
0 likes · 7 min read

The article details Meituan's deployment of large-scale heterogeneous graph recall for in‑store recommendation ads, covering full‑scene graph construction, graph pruning, dynamic negative sampling, spatiotemporal sub‑graph fusion, and performance optimizations that together raise offline hit‑rate by over 5% and online revenue per search by 10‑15%.
Graph Neural NetworksMeituanRecommendation System
0 likes · 25 min read

The presentation by Tencent expert Yuan Yi details the Wuliang deep learning system for recommendation, covering its background, technical challenges such as massive data and real‑time requirements, the parameter‑server based solutions for training and inference, model compression techniques, and continuous online deployment strategies.
deep learninglarge-scale trainingparameter server
0 likes · 14 min read

Alibaba’s advertising AI team presents a comprehensive study of four new multimodal copywriting tasks—image overlay text generation, video narration, text style transfer, and detail-page extraction—detailing model architectures, training on billions of images, experimental results, and practical deployment in the “Xiyu” product.
advertising copydetail extractionlarge-scale training
0 likes · 17 min read

This article introduces Baidu's PGL2.2 graph learning platform, explains graph modeling and message‑passing GNN techniques, details training strategies for small, medium and large graphs, showcases node classification and link‑prediction methods, and describes how the framework is applied in search, recommendation, risk control, and knowledge‑graph competitions.
Graph Neural NetworksPGL2.2knowledge graphs
0 likes · 15 min read

This article describes how Ctrip improved hotel recommendation by iterating from logistic regression to GBDT and deep neural networks, designing continuous and discrete features, adopting multi‑task learning with click and conversion signals, and building a large‑scale distributed DNN training and unified feature‑processing framework to boost model accuracy and engineering efficiency.
CtripDNNfeature engineering
0 likes · 15 min read

This article recounts Alibaba‑Mama researcher Jing Shi’s presentation on the evolution of deep learning for click‑through‑rate (CTR) estimation, covering the shift from handcrafted features and linear models to piecewise linear MLR, end‑to‑end neural networks, multi‑interest user modeling, and large‑scale distributed training challenges.
AdvertisingCTR predictiondeep learning
0 likes · 16 min read
