#GPU memory bandwidth optimization
211 related articles
·3 min
Kimi Delta Attention Explained: The Evolution from Standard Attention to Linear Attention
A deep dive into Kimi Delta Attention (KDA): tracing the evolution from quadratic Softmax attention through linear attention, Delta rules, and gated decay mechanisms, with insights on associative memory and hardware optimization.
Read more →
·3 min
Kimi Delta Attention Explained: The Evolution from Standard Attention to Linear Attention
Deep dive into Kimi Delta Attention (KDA): from standard Softmax attention's quadratic bottleneck through linear attention, Delta Rule, and gated decay mechanisms — the complete evolution explained.
Read more →
·3 min
ncnn Vulkan Cross-Platform Edge Inference in Practice: Solving the GPU Fragmentation Deployment Challenge
Facing GPU fragmentation on edge devices, the PostSlate team used ncnn's Vulkan backend for cross-platform ML inference, achieving 10× speedup on RTX 4070 with half the model size and zero runtime installation.
Read more →
·3 min
Kimi K3 Launch: A 2.8 Trillion Parameter Flagship Model Makes Its Grand Entrance
Moonshot AI launches Kimi K3 with 2.8 trillion parameters and 1M token context. Google delays Gemini 3.5 Pro, AI coding tools upgrade collectively as competition shifts to coding and Agent capabilities.
Read more →
·5 min
July 24 AI Daily Brief: Flux 3 Multimodal Release, Kimi K3 Lags in Testing, Qwen Tops TTS Rankings
July 24 AI news: Black Forest Labs launches Flux 3 multimodal model, Kimi K3 lags in US-UK gov tests, Alibaba Qwen tops TTS rankings, Etched raises $300M, AMD unveils MI430X.
Read more →
·4 min
Kimi K3 Released: 2.8 Trillion Parameter Open-Source Model Tops Global Coding Leaderboard
Moonshot AI releases Kimi K3, a 2.8 trillion parameter open-source model using MoE architecture that tops the global frontend coding arena at under $1 per task, beating GPT and Claude.
Read more →
·4 min
Colibri: Running a 744-Billion-Parameter LLM on an Ordinary Laptop
Colibri uses MoE hot-cold separation and 4-bit quantization to run 744B-parameter models like GLM 5.2 on consumer hardware. Learn about its three-tier memory architecture and speculative decoding.
Read more →
·4 min
Laguna S 2.1 Deep Dive: Hands-On Local Deployment of a 118B Parameter Coding Model
In-depth review of Poolside's Laguna S 2.1 open-source coding model: MoE architecture, RL training, DGX Spark local deployment, and real-world agentic coding tests with 8B active parameters.
Read more →
·3 min
Samsung Z Fold 8 Hands-On: The Compact Foldable Is the Biggest Surprise
Samsung unveils the Z Fold 8, Z Fold 8 Ultra, and Z Flip 8. The Z Fold 8's passport-style design weighs just 201g with a dramatically improved crease, preemptively countering Apple's rumored foldable iPhone.
Read more →
·3 min
Samsung Z Fold 8 Hands-On: The Compact Foldable Is the Biggest Surprise
Samsung launches Z Fold 8, Z Fold 8 Ultra, and Z Flip 8. The Z Fold 8's passport-style design weighs just 201g with dramatically reduced crease, preemptively countering Apple's rumored foldable iPhone.
Read more →
·2 min
Chrome Built-in AI Explained: How Locally Running Models Protect Your Privacy
Deep dive into Chrome Built-in AI technology, exploring how running AI models locally in the browser achieves zero data uploads, instant responses, and stronger privacy protection.
Read more →
·2 min
Chrome Built-in AI Explained: How Locally Running Models Protect Your Privacy
Explore Chrome Built-in AI technology and how running AI models locally in the browser enables zero data upload, instant responses, and stronger privacy protection.
Read more →
·3 min
Apple Silicon LLM Inference Speed Benchmarks: A Complete Guide to Local Large Model Performance on Mac
In-depth analysis of Apple Silicon local LLM inference speed benchmarks covering M-series memory bandwidth, model quantization, MLX framework optimization, and Mac configuration guidance.
Read more →
·3 min
Monocular RGB Human Mesh Reconstruction: Real-Time Motion Capture at 18 FPS with Rust + CUDA
Deep dive into a real-time 3D human mesh reconstruction project using a single RGB camera, built with Rust, Candle, and CUDA, achieving 55ms/frame on RTX 5080. Exploring its architecture, Metal porting plans, and applications in VTuber, AR/VR, and sports analysis.
Read more →
·3 min
Monocular RGB Human Mesh Reconstruction: Real-Time Motion Capture at 18 FPS with Rust + CUDA
A deep dive into a real-time 3D human mesh reconstruction project using a single RGB camera, built with Rust, Candle, and CUDA, achieving 55ms/frame on an RTX 5080.
Read more →
·3 min
AMD CDNA5 Architecture Deep Dive: Technical Evolution and the AI Computing Competition Landscape
Deep analysis of AMD's CDNA5 architecture covering Chiplet packaging upgrades, HBM memory evolution, and low-precision compute optimization, examining how AMD challenges NVIDIA's AI chip dominance.
Read more →
·3 min
Kimi K3 Launches on Ollama Cloud: Free Tier, User Experience, and Developer Guide
Kimi K3 officially launches on Ollama Cloud as an "extra high usage" model. This guide covers free tier quotas, cloud inference experience, technical advantages, and how developers can seamlessly call this high-performance LLM.
Read more →
·2 min
DeepSeek V4 Official Release Imminent: What 1.6 Trillion Parameters Really Means
DeepSeek V4 official release is imminent. V4 Pro features 1.6 trillion parameters with million-character context. Explore the D-SPARK acceleration framework, dynamic pricing, and V4's leap to production-grade AI.
Read more →
·4 min
The Dao of DeepSeek: Liang Wenfeng on Vision, Restraint, and the Philosophy of Open Source
Analysis of DeepSeek founder Liang Wenfeng's rare investor dialogue, exploring the company's vision-driven culture, strategic restraint toward AGI, and open-source philosophy in the US-China AI race.
Read more →
·2 min
Google Flash and Flash-Lite Models Upgraded and Live: A New Lightweight, High-Efficiency AI Option
Google rolls out upgraded 3.6 Flash and 3.5 Flash-Lite models. Learn about the Flash series' positioning, upgrade highlights, and value for developers.
Read more →