RAM vs VRAM for Local LLMs: How CPU Offload Works and What It Costs in Speed
RAM vs VRAM for local LLMs: why memory bandwidth sets speed, how CPU offload splits a model, worked speed ceilings for a 70B model, and when more RAM helps.
-
-
ComfyUI Hardware Requirements: How Much VRAM and RAM You Need for Image Generation
ComfyUI hardware requirements for SD 1.5, SDXL, SD 3.5 and Flux: model file sizes, how much VRAM and RAM to plan for, GPU choices with reference prices.
-
Mini PC for Large LLMs: Ryzen AI Max+ 395, DGX Spark and Mac with 128GB of Unified Memory
Mini PC for large LLMs: Ryzen AI Max+ 395, DGX Spark and Mac Studio compared on memory, bandwidth and reference prices, and why big memory does not mean fast.
-
Used GPU for Local LLMs: How to Pick and Test an RTX 3090 or 4060 Ti 16GB
Used GPU for local LLMs: price per GB of VRAM for the RTX 3090 and 4060 Ti 16GB, an inspection checklist, burn-in tests, and and the listings that burn buyers.
-
12GB vs 16GB VRAM for Local LLMs: What Runs, and RTX 3060, 4060 Ti or 5060 Ti
12GB vs 16GB VRAM for local LLMs: which models fit, and how the RTX 3060 12GB, 4060 Ti 16GB and 5060 Ti 16GB compare on price and memory bandwidth.
-
AMD GPU for Local LLMs: ROCm vs Vulkan, and Can the RX 7900 XTX or 9070 XT Do the Job?
Using an AMD GPU for local LLMs? See Ollama’s support list, choose ROCm or Vulkan, compare RX 7900 XTX and 9070 XT bandwidth, and fix common errors.
-
24GB VRAM PC Build: A Used RTX 3090 Machine at About 14299 Yuan, and How It Compares With an RTX 4090
24GB VRAM PC build: a used RTX 3090 machine at about 14299 yuan against about 33214 for a 4090, with parts, models that fit, power and cooling, and risks.
-
Entry-Level PC Build for Local LLMs: A Roughly 8000 Yuan RTX 3060 12GB Machine
Entry-level PC build for local LLMs: an RTX 3060 12GB and i5-12600KF machine at about 7899 yuan, with prices, models it runs, upgrades and where not to skimp.
-
LLM Context Length VRAM: How the KV Cache Works, Formulas, Worked Examples and Fixes
LLM context length VRAM guide: the KV cache formula, worked examples for 7B to 70B models, f16 vs q8_0 vs q4_0 cache, Ollama defaults and out-of-memory fixes.
-
Dual GPU LLM Build: Two Used RTX 3090s for 48GB, Parts, Power, Cooling and Limits
Dual GPU LLM build guide: two used RTX 3090s for 48GB, what fits, PCIe, PSU, cooling, llama.cpp and Ollama splitting, derived cost, and when one card is better.
