Local AI Data Center on Salvaged Hardware

Built a local AI Data Center from salvaged DDR3-era hardware capable of running large language models up to 122B parameters. The system runs on a single Huananzhi X79 dual-socket motherboard with a custom-rewritten BIOS, powering a Vega 64 + 4x Radeon Instinct MI25 16GB GPU array (80GB total VRAM) through ROCm.

Everything in this build was salvaged or pulled from broken machines:

Motherboard: Huananzhi X79 dual-socket (Intel C602 chipset) with a fully rewritten BIOS to fix the factory's broken PCI lane allocation from a copy-pasted 4-socket server template.

GPU array: 1x AMD Vega 64 + 4x AMD Radeon Instinct MI25 16GB, totaling 80GB VRAM. All Vega architecture (GCN 5), synchronized through custom PCI configuration.

Memory: DDR3 platform. The entire Data Center runs on hardware most people would consider obsolete.

Compute stack: ROCm for GPU compute, running on a custom-optimized Linux kernel.

The system handles quantized large language models and Mixture of Experts architectures:

Qwen 3.5 122B: The largest model in the stack, running quantized across 80GB VRAM.

Llama 3.1 70B: Meta's flagship open model.

Gemma 4 32B Coder: Google's coding-specialized model.

Alpha Agent 80B: A custom-built model, developed and trained in-house.

Quantization and MoE routing allow these models to fit and perform on hardware that, on paper, has no business running them.

This Data Center exists because I refuse to let corporations profit at my expense while I survive on less than a living wage. I've been through 4 heart attacks and 2 hours of clinical death. My government classifies me as disabled and pays me below the minimum cost of living.

So instead of renting compute from companies that charge more than I earn, I built my own. From salvaged parts, broken machines, and hardware nobody else wanted. Every component was hand-optimized: BIOS rewritten from scratch, kernel ported and tuned, three generations of GPUs forced into synchronized operation. 80GB of VRAM running models that are supposed to require enterprise-grade infrastructure.

The bottleneck is rarely the hardware. It's how well you understand it. I don't follow the rules. I bend them to my will.