Simplifying Transformer Blocks

Simplifying Transformer blocks by removing skip connections, projections and normalization; introduces Simplified Attention to reduce parameters and raise training throughput.

Mamba: New Selective State Space Model vs Transformer

Mamba: a selective SSM state-space model that generalizes S4 to enable linear long-context scaling, million-token sequences, and improved language modeling.

Building a Multilingual Open-Ended QA Dataset

OMGEval presents an open-source multilingual open-ended QA benchmark (804 Chinese prompts) localized from AlpacaEval, using Text-Davinci-003 baseline and GPT-4 evaluation.

Memory Interface Chips for AI Servers

Technical overview of AI server interconnects and components: DGX H100 architecture, PCIe switches and Retimers, and DDR5 memory interface chip trends.

What Is AI Computation and How Does It Work?

AI overview with latent space representations and practical applications in manufacturing and semiconductor manufacturing, including predictive maintenance and quality assurance.

GPT and Neural Networks: Underlying Mechanisms

Technical overview of neural networks and GPT: how images and text are vectorized, forward/backpropagation, gradient descent training, activations, and prediction.

Latest Trends in Machine Learning Hardware Performance

Analysis of ML hardware trends across GPUs and accelerators, quantifying compute performance, interconnects, cost-performance, and energy efficiency.

What foundational devices do AI chips require?

Tsinghua's Future Chip Forum recap: Wei Shaojun outlines constraints for zettascale systems, device needs and prospects for 3D integration.

Challenges for AI Chips in the Large-Model Era

Analysis of heterogeneous computing and AI chips in the large-model era: performance gaps, CUDA ecosystem limits, pooled training, and evaluation needs.

Common Deep Learning Applications in AI

Survey of deep learning applications in AI: image recognition, NLP, speech, recommendation systems, autonomous driving, healthcare, cybersecurity, and VR.

AI's Impact on Industrial Software

Siemens Digital Industries white paper on AI in electronic systems design, exploring AI methods and applications for PCB design, component selection, layout and verification.

CNN Image Classification Strategy

Explains how BagNets show ImageNet classification relies on local bag-of-features strategies, revealing CNN texture bias, patch-based evidence and robustness issues.

Explaining Convolution in Simple Terms

Explains the meaning of convolution—why we flip (fold) and multiply—using signal analysis, dice probability, and image processing kernels as examples.

AI Compute Server Technology and Industry Chain Report

Technical overview of compute scheduling in the compute network, covering orchestration, cross-domain scheduling, cost and latency tradeoffs, and platform architecture.

Change Detection in Remote Sensing Using Foundation Models

Time-Traveling Pixels integrates SAM into remote sensing change detection, using low-rank fine-tuning and a Time-Travel Activation Gate to mitigate spatial-semantic domain shift.

ORB-SLAM3 Architecture and Workflow Analysis

Overview of ORB-SLAM3 architecture and visual-inertial SLAM: tracking, local mapping, loop/map merging, Atlas and IMU-camera fusion for pose estimation and optimization.

Convolutional Neural Networks: Analysis and Practice

Overview of convolutional neural networks: how filters, sliding-window local matching, convolution, ReLU activations, and pooling produce feature maps for image classification.

Monocular Vision Methods for Vehicle Distance Estimation

Review of monocular ranging algorithms and imaging geometry for forward collision warning, covering camera pose, lane-width distance estimation and accuracy metrics.

Will Gemini Disrupt the Large Model Market?

Technical overview of Google Gemini, a multimodal foundation model family (Ultra, Pro, Nano), its benchmarks vs GPT-4, multimodal capabilities, and TPU efficiency.

How Compute Chips Deliver Performance for ChatGPT

Explains how PCIe and compute cards form the compute foundation for generative AI systems, covering bus roles, testing, reliability, and high-speed interconnects.