The original title is "From API to GPU, Week 1: Understanding NVIDIA DGX Spark Environment"

Short summary

A software engineer begins a 32-week journey from AI API consumer to local LLM systems architect, starting with standing up and inventorying an NVIDIA DGX Spark machine. The post covers CPU vs GPU architecture fundamentals, walks through real nvidia-smi output field-by-field, and explains how the Spark's unified 121 GiB memory pool differs from traditional discrete GPU VRAM. A companion bash script automates the full hardware inventory via SSH, making every claim verifiable by command output.

• •DGX Spark is an aarch64 ARM64 machine with a 20-core CPU and NVIDIA GB10 GPU using unified 121 GiB memory instead of separate VRAM
• •nvidia-smi serves as the primary GPU monitoring tool, reporting driver version, CUDA support, temperature, power, and utilization
• •A companion inventory script reproduces all commands via SSH so readers can verify every hardware claim themselves

Generated with AI, which can make mistakes.