Build Your Own AI server

I built a DIY AI server to see how far a home setup can go without a DGX or a pricey custom workstation. This video covers the options, the parts I chose (and why), and real benchmarks across LLMs, image generation, and ML training.

Chapters (timestamps):
00:00 Intro
00:54 AI server options
01:29 Parts: Motherboard
02:24 Parts: CPU
02:48 Parts: RAM
02:56 Parts: Cooling
03:18 Parts: PSU
03:28 Parts: GPU
04:33 LLM Benchmarks: gpt‑oss 20B (parallel)
06:04 LLM Benchmarks: gpt‑oss 20B (single‑request)
07:28 LLM Benchmarks: gpt‑oss 120B (sharded)
08:20 Image Gen Benchmarks: FLUX.1 dev
09:48 ML Training Benchmarks: ResNet50
11:19 Cost comparison
12:09 Wrap‑up & plans
14:18 Extra stuff / credits

Corrections:
08:22 FLUX.1 isn’t a Stable Diffusion model but it’s a transformer-based, rectified-flow diffusion model.
10:18 not a GPU, but this is Nvidia's Grace CPU. Wrong image.

Related Videos:
   • How NOT To Cook Your GPUs  
Deep Dives with Christina —    • 01 - Choosing an NVIDIA GPU for AI 2025  

Ways to support:
Subscribe -- https://www.youtube.com/c/JunYamog?su...

Videos Mentioned:
‪@theNittyGritty‬ — Why I am not gonna buy the NVIDIA DGX Spark    • Why I am not gonna buy the NVIDIA DGX Spark  
‪@DigitalSpaceport‬ — INSANE Home AI Server - Quad 3090 Build    • INSANE Home AI Server - Quad 3090 Build  

References:
gpt-oss - https://openai.com/index/introducing-...
vLLM - https://docs.vllm.ai/en/latest/
llama.cpp - https://github.com/ggml-org/llama.cpp
nVidia GenAI-Perf - https://docs.nvidia.com/deeplearning/...
Apple MLX - https://opensource.apple.com/projects...
ML Training ResNet50 (forked from M Bourke's benchmarks and added torch discributed for multi GPU runs) - https://github.com/jkyamog/mac-ml-spe...

If you’d rather buy than build, check out Bizon Tech (not sponsored) — https://bizon-tech.com/deep-learning-...

Motion 5 animations by Simple Video Making (not sponsored) — https://simplevideomaking.com/