A practical guide to selecting hardware for running local LLMs in embedded software development workflows. Covers memory sizing math for model weights at different quantization levels (FP16, INT8, INT4), KV cache overhead, and MoE architecture memory requirements. Compares discrete GPU VRAM (NVIDIA RTX 5090 with 32 GB) against unified memory systems (Framework Desktop with Ryzen AI Max+, Mac Studio, NVIDIA DGX Spark) across capacity, bandwidth, and cost dimensions. Argues that for embedded development tasks like driver review and agentic codebase analysis, memory capacity matters more than token speed, making unified memory the economically sensible choice for large models. The author settled on a dual NVIDIA DGX Spark setup (256 GB combined unified memory) for its CUDA-native ecosystem compatibility.
