-
AI server 3090
May 2026 picks: 2x RTX 3090 (48GB) for the dense-model workhorse; 2x RTX 5060 Ti 16GB (32GB) for budget MoE with --cpu-moe; 2x RTX 2080 Ti 22GB modded for value (Qwen 3. 6 27B at 38 tok/s); 1x 3090 + 1x 4090 for mixed-card pipeline parallelism. cpp and Ollama handle. Want to build a GPU home server for running quantized models? Here's some tips and tricks for setting up the server. RTX 3090: Two RTX 3090s with NVLink are a common choice for running large AI models. Previously I have built one but only for mining where those GPUs were connected via PCIE x1 risers. 0 x16 so the thing looks slightly different. Building a full DIY rig is a high base cost with inflation with every new recent dual slot capable motherboard checking in above $100. AI from The Basement: My latest side project, a dedicated LLM server powered by 8x RTX 3090 Graphic Cards, boasting a total of 192GB of VRAM. This blogpost was originally posted on my LinkedIn profile in July 2024. Backstory: Sometime in. A 70B model that can't fit on one 24GB card runs at 16-21 tok/s across dual RTX 3090s. You need server-grade platforms.
[PDF Version]