💡AI Data Center Built Using E-Waste GPUs
Building an AI server on a shoestring budget
TL;DR
A developer built an AI data center using AMD V620 GPUs from e-waste, Intel Core i9 10900X CPU, and salvaged RAM & SSDs. The setup runs Ubuntu 24.04 and supports llama.cpp inference server.
Developer builds a DIY AI data center with AMD's unsold V620 GPUs from e-waste, Intel Core i9 10900X CPU, and salvaged RAM & SSDs. The build runs Ubuntu 24.04 and supports llama.cpp inference server for AI models like Gemma4. This setup is a cost-effective solution for those looking to experiment with large language models without breaking the bank. V620 GPUs come with 32GB of fast VRAM, making them ideal for parallel matrix math tasks.
Key Points
AMD V620 GPU: 32GB fast VRAM, never sold to public due to overproduction and lack of software support
Intel Core i9 10900X CPU: Considered the worst Intel produced in last twenty years
Motherboard X299 platform with four PCIe x16 slots for GPU installation
Power supply: 1600W to handle power requirements of four GPUs and inefficient processor
Fan controller built using Arduino Nano clone controls fan speed based on GPU temperature
Why It Matters
If you're experimenting with large language models but have a tight budget, this setup shows how to build an AI data center for under $1000. The AMD V620 GPUs offer impressive performance at a fraction of the cost compared to new hardware. However, it requires significant DIY skills and patience to get everything working.
Frequently Asked Questions
Why does this matter?
If you're experimenting with large language models but have a tight budget, this setup shows how to build an AI data center for under $1000. The AMD V620 GPUs offer impressive performance at a fraction of the cost compared to new hardware. However, it requires significant DIY skills and patience to get everything working.
What happened?
A developer built an AI data center using AMD V620 GPUs from e-waste, Intel Core i9 10900X CPU, and salvaged RAM & SSDs. The setup runs Ubuntu 24.04 and supports llama.cpp inference server.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 2,975 builders reading daily.