ALLPCB
MT8391 (Genio 720) Specifications for Edge AIoT
Overview of the MediaTek MT8391 (Genio 720) edge AI platform: 6 nm octa-core CPU, 10 TOPS NPU, dual ISPs, LPDDR5 support and multi-interface connectivity for AIoT devices.
Deep Learning for Defect Detection
Explore deep learning for defect detection in industries, offering accurate solutions for quality control with advanced frameworks.
Impact of Irrelevant Retrieved Passages on LLM Outputs in RAG
Retrieval-augmented generation robustness analysis: semantically related but answer-irrelevant retrieved fragments and higher fragment counts degrade LLM accuracy and confidence.
Sensors Enable AI to Perceive the World
Edge AI and vision sensors overview: embedded ML platforms, real-time image processing, depth sensing, and on-device inference for industrial, smart city, and IoT applications.
Parallel Computing and Acceleration Techniques for Neural Networks
Overview of parallel computing and acceleration for neural networks, covering data/model parallelism, GPU/TPU and software optimizations like mixed precision.
Classical Computer Vision vs. Deep Learning
Analysis of deep learning in computer vision: strengths, limits, dataset biases, comparison with classical vision methods, interpretability and risks in safety-critical applications.
What is an AI Edge Compute Box?
Comprehensive overview of an AI edge compute box: definition, operation, features, security, and industry use cases for real-time edge computing and computer vision.
Infrared Thermal Imaging in Anti-Drowning Systems
Analysis of infrared thermal imaging for anti-drowning systems: effective for 24/7 perimeter detection in prohibited areas but unsuitable for continuous swimmer tracking.
Thermal and EMI Challenges in AI Chip Design
Explore thermal and EMI challenges in AI chip design, focusing on heat dissipation and noise suppression for high-performance computing.
MambaQuant: Post-Training Quantization for Mamba Models
MambaQuant: PTQ for Mamba models using KLT-enhanced and smoothed fused rotations to enable high-accuracy W8A8/W4A8 post-training quantization with <1% loss.