#PCIe bandwidth

56 related articles

Jul 29, 2026·3 min

Local LLM Agent Lagging? Lowering num_gpu Is the Key

Local LLM crashing in Agent frameworks? The issue may be num_gpu set too high. Learn what num_gpu really controls (GPU layer offloading, not GPU count) and how to tune it for stable Agent performance.

Jul 28, 2026·4 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.

Jul 28, 2026·4 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.

Jul 28, 2026·4 min

Framework 13 Pro In-Depth Review: The Modular Laptop That Finally Stops Compromising

Framework 13 Pro in-depth review covering build quality, display, performance, and Linux experience. This modular laptop achieves 85% of MacBook Pro's refinement with far superior repairability and upgradeability.

Jul 28, 2026·3 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.

Jul 28, 2026·3 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.

Jul 28, 2026·3 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.

Jul 25, 2026·6 min

How the Creator of Redis Squeezed a 284B-Parameter DeepSeek into an Ordinary Laptop

Redis creator antirez built DS4 "Dwarf Star," a pure-C inference engine, using asymmetric quantization to compress DeepSeek V4 Flash from 500GB to 80.8GB for local 128GB unified-memory deployment at 26.7 tokens/sec.

Jul 25, 2026·6 min

How the Creator of Redis Squeezed a 284-Billion-Parameter DeepSeek Into an Ordinary Laptop

Redis author antirez built the pure-C inference engine DS4 "Dwarf Star," compressing DeepSeek V4 Flash from 500GB to 80.8GB via asymmetric quantization for local deployment on 128GB unified memory at 26.7 tokens/sec.

Jul 23, 2026·7 min

Alibaba Open-Sources 2.4T Qwen Model as Chinese AI Sweeps the Global Competitive Landscape

Alibaba open-sources a 2.4 trillion parameter Qwen model and launches the Qwen Token Plan. Chinese models surge, Kimi K3 tops global rankings, and China's AI is reshaping the global competitive landscape.

Jul 22, 2026·3 min

V100 GPU Showdown: 4×16G PCIe vs. 2×32G SXM — Which Gives You More Bang for Your Buck?

Benchmarking 4×V100 16G PCIe vs. 2×V100 32G SXM adapter for local LLM inference. Prefill speed, decode speed, power limits, and bandwidth bottlenecks analyzed.

Jul 21, 2026·2 min

MiniMax M3 Local Deployment: 23B Active Parameters ≠ 23GB VRAM — A Complete Guide to MoE Architecture

MiniMax M3 is a 428B MoE model. Its 23B active parameters reflect per-token compute, not VRAM needs. Learn the MoE trade-offs, quantization options, and deployment paths to avoid the most common misconception.

AirLLM: How a 4GB GPU Can Run a 70B Mo…

Jul 19, 2026·2 min

AirLLM: How a 4GB GPU Can Run a 70B Model — Core Principles and Use Cases

AirLLM is an open-source project that uses layer-by-layer inference to run 70B LLMs on a single 4GB GPU. Learn how it works, its tradeoffs, and ideal use cases.

Apple M7 Ultra Leaked: Can 1.5TB Unifi…

Jul 16, 2026·2 min

Apple M7 Ultra Leaked: Can 1.5TB Unified Memory Challenge NVIDIA Blackwell?

Apple's M7 Ultra chip rumored to feature 1.5TB unified memory targeting NVIDIA Blackwell-level AI performance. A deep dive into UMA advantages, what "matching Blackwell" really means, and Apple's AI strategy.

Jul 16, 2026·2 min

How Large a Local LLM Can You Run on a Laptop with 8GB VRAM? A Configuration Guide

Complete guide to running local LLMs on a laptop with 8GB VRAM: real usable memory, quantization estimation, Q4 7B/8B model recommendations, Ollama setup, GPU offloading, and agent development tips.

Apple M7 Ultra Chip Leaked: Can 1.5TB …

Jul 14, 2026·2 min

Apple M7 Ultra Chip Leaked: Can 1.5TB Unified Memory Reshape Local AI Inference?

Reddit leaks suggest Apple's M7 Ultra chip could feature up to 1.5TB unified memory. We analyze the architecture, pricing debate, bandwidth limits, and ecosystem trade-offs for local LLM inference.

Building an eGPU for Local LLM on a $1…

Jul 14, 2026·2 min

Building an eGPU for Local LLM on a $1,000 Budget: A Practical Buying and Setup Guide

Building an eGPU for local LLMs on a $1,000 budget? This guide covers GPU selection, dock costs, RTX 3090 vs 3060 value, and top tools like Ollama and llama.cpp.

JAX Host Offloading: A Practical Guide…

Jul 14, 2026·2 min

JAX Host Offloading: A Practical Guide to Breaking Through LLM Training Memory Bottlenecks

Memory capacity is the core bottleneck in LLM training. This guide explores JAX-based host offloading — covering optimizer state offloading, activation strategies, PCIe bandwidth trade-offs, and how it complements activation recomputation.

Jul 13, 2026·2 min

Local Image-to-3D on Apple Silicon: 20 Seconds, 2GB RAM — Modelr Open-Source App Tested

A developer ports Tencent's Hunyuan3D to Apple MLX. M4 Max benchmarks: shape generation in 20s, 5.6GB peak RAM. Quantized versions run on iPhone. Modelr is the first local image-to-3D app for Apple Silicon.

How Fixing Three Bugs Made Qwen3.5-122…

Jul 13, 2026·3 min

How Fixing Three Bugs Made Qwen3.5-122B Run Reliably on Mac Studio

A developer fixed three critical bugs to make Qwen3.5-122B run reliably as a daily driver on Mac Studio. Explore memory management, inference stability, and quantization precision.