I Built The Smallest & Affordable AI Computer — for $150

Tech concept / project overview

• A DIY “ultra-slim” AI computer built for ~$150 that can run offline AI workloads:
  • LLMs (Llama via Ollama)
  • Vision-language models (LLaVA / “Lava”)
  • Image generation (Stable Diffusion)
  • Mentions video generation (~5 seconds) from Stable Diffusion workflows.
• Form factor: about ~24 mm thick, slimmer than typical mini PCs / Jetson-style devices.
• Core idea: create a compact “robot brain” (referenced as future use in a robot project).

Hardware components and target specs

• Goal: stay within ≤ $150 while running local LLMs.
• Selected parts:
  • Zema Blade: compact quad-core x86 SBC, ~16 GB RAM, ~$50 (marketplace)
  • Tesla P4 GPU: data center card with 8 GB dedicated VRAM, ~$60 used (eBay)
• Intended capability: run larger models locally (claimed up to ~13B class) due to dedicated VRAM rather than shared memory.

Major engineering problems + solutions (as described)

1. Power compatibility issue

   • Problem: the Tesla P4 wouldn’t power from the Zema Blade’s power delivery.
   • Solution: use a PCI riser with separate power input:
     • power the Zema Blade separately
     • power the Tesla P4 separately
     • use a cleaner cabling/adapter approach for integration.

2. Heat management / thermal runaway risk

   • Problem: initial testing (Ubuntu + Ollama + LLM) caused the system to freeze.
   • Observation: GPU temps reportedly spiked to ~90–95°C.
   • Root cause: the Tesla P4 has no fan, so passive airflow was insufficient.

3. Mechanical constraint: coin-slim enclosure

   • Constraint: enclosure width/thickness needed to be ~24 mm, under a ~26 mm target.
   • Solution: redesign the enclosure as part of the cooling system:
     • Fusion 360 CAD airflow/vent design
     • Aluminum base as a passive heat sink
     • 3D-printed fan shroud to direct air over the Tesla GPU fins
     • blower fan mounting inside the slim case
     • XT60 connector and integrated power switch for one-supply control.

4. Board-level thermal interface

   • The Zema Blade case was removed to reduce height (its body previously acted as part of the heat sinking).
   • Added a 2 mm copper shim + thermal paste between the Zema Blade CPU and the aluminum base, so the case/base functions as the heat sink.

Power system details

• Added a single slim ~200 W PCU (power supply) sized for the build.
• Peak claimed draw: ~120–150 W, leaving headroom for the PSU.
• Wiring approach:
  • PSU output to a hub/connector powering:
    • Zema Blade
    • Tesla P4 via riser
    • the blower fan
  • Uses XT60 for power + switch integration.

Software / workload testing results

LLM test

• Ran Llama 13B via Ollama.
• Result: worked smoothly with no crash.
• Temperature improvement:
  • previously reported ~90–95°C
  • down to about ~39°C in this setup
• Also tested 8B, reporting faster/usable performance.

Vision-language test (offline)

• Ran LLaVA (“Lava”) on an input image.
• Claimed offline/no-internet operation.
• Reported temperature stable around ~48°C.

Generative media test

• Ran Stable Diffusion for image generation.
• Also generated a ~5-second video (as claimed in subtitles).
• All performed offline on the same compact build.

Review / comparison analysis vs common options

• Compared against:
  • Jetson Nano Super (~$500)
    • Claimed limitation: 8 GB shared memory roughly caps it at ~7B parameter models (claim).
    • Efficiency and robotics friendliness are praised, but not model-size capability.
  • Thunder Robot “Mix” (RTX 4060 mini PC, >$1,200)
    • Very powerful but expensive; questioned for suitability with compact form factors.
• Claimed advantage of the ~$150 build:
  • Tesla P4 + 8 GB dedicated VRAM enabling 13B class runs at lower cost.
• Caveat acknowledged:
  • Not as power-efficient as Jetson, but viewed as better cost/performance.

Tutorials / guides implied by the build narration

• The narration implies a step-by-step approach:
  1. Choose an SBC + used GPU to gain VRAM capacity
  2. Handle PCI riser + external GPU power
  3. Design a thermal/mechanical enclosure (CAD + 3D printed parts)
  4. Use a copper shim + thermal paste to couple board heat to the case
  5. Consolidate wiring/power using XT60
  6. Validate with stress tests and progressively heavier workloads (LLM → vision → diffusion/video)

Main speakers / sources

• Speaker: the video narrator/builder (self-described creator who built the device “myself,” running tests and comparisons).
• Referenced products/tools: Zema Blade, Tesla P4, Ubuntu, Ollama, LLaVA (“Lava”), Stable Diffusion, Fusion 360, Jetson Nano Super, and “Thunder Robot Mix.”