Shared first on X by David Hendrickson (@TeksEdge). 2x RTX 3090 GPUs linked with an NVLink bridge, giving 48GB of pooled VRAM via tensor parallel (TP=2). A ~$4K build that runs 70B+ models at Q4/Q5 quantization using vLLM, llama.cpp, or ExLlamaV2.
Original post: https://x.com/TeksEdge/status/2047479832502743444