Ultra-Budget Local LLM Build: Dual GTX 1660 + Intel Xeon E5-2650 v4
Building a high-throughput 12GB VRAM local AI server for under $180 total using enterprise scrap server hardware and dual Turing GPUs.
YouTube Companion Video
Watch the full step-by-step physical assembly, BIOS tweaks, and live terminal testing.
1. The Core Engineering Challenge
Running 7B and 8B parameter models like Meta Llama 3.1 8B or Mistral 7B locally requires at least 8GB to 10GB of VRAM for 4-bit and 5-bit quantization with an 8k context window. Consumer advice usually funnels creators into buying a single RTX 3060 12GB (\$280+) or an RTX 4060 Ti 16GB (\$450+).
In the Roswellien lab, we looked at this from an enterprise perspective: Why not pair two second-hand Turing cards (GTX 1660 Super 6GB) on an enterprise server socket?
✓ Why Xeon E5-2650 v4?
Provides 40 native PCIe 3.0 lanes. Consumer AM4 / LGA1700 platforms drop your second GPU slot to x4 or x8 through the chipset. Broadwell-EP runs both 1660s at full bandwidth simultaneously.
✓ Why Turing FP16 Support?
Unlike older Pascal cards (GTX 1060 / 1070) which suffer severe performance penalties on FP16 matrix math, the 1660 series has dedicated FP16 execution units, running quantized GGUF and AWQ layers with full speed.
2. Bill of Materials (Budget Breakdown)
| Component | Target Cost | Primary Source | Affiliate Link |
|---|---|---|---|
| Intel Xeon E5-2650 v4 (12C / 24T) | $14 | AliExpress / eBay | View Part → |
| 2x NVIDIA GeForce GTX 1660 Super (6GB x 2) | $110 ($55 ea) | eBay Used | View Part → |
| Machinist X99 Dual-Slot Motherboard | $32 | AliExpress | View Part → |
| 16GB DDR4 ECC Registered RAM (Quad Channel) | $16 | eBay Server Pull | View Part → |
*Prices reflect average second-hand eBay / AliExpress pricing in late 2024 - 2026.
3. Real-World LLM Generation Benchmarks
At 18 to 21 tokens per second, interactive chat and coding assistance feels instantaneous. For under \$180, this completely matches or exceeds the real-world inference speed of expensive off-the-shelf workstation cards.
4. Software Configuration & Execution
A. Running Ollama Across Both GPUs
Ollama automatically detects multiple CUDA devices and distributes model layers across both cards evenly:
CUDA_VISIBLE_DEVICES=0,1 ollama run llama3.1:8bB. llama.cpp Tensor Splitting
For maximum control over context buffer allocation, use llama-server with exact tensor split ratios:
./llama-server -m Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf --tensor-split 1,1 -ngl 99 -c 8192Have questions about this build?
Drop a comment on the YouTube transmission or check out our full hardware catalog in the Gear Lab.
Explore All Lab Gear →