Decommissioned NVIDIA enterprise GPUs—K80, P100‑16GB, V100‑16GB—remain a cheap source of idle VRAM, selling for $60, $75, and under $200. The author spent a winter running benchmarks on these cards to gauge their value in modern workloads. The goal is an inexpensive 4U GPU node for a homelab, fitting three GPUs and a 10GB NIC into a standard ATX or rack‑mount case.

A critical note: all hardware examined is EOL, lacking future CUDA updates and being less power efficient, yet the author argues homelab use is still viable with older software like llama.cpp and Docker containers. The benchmark suite, published on GitHub, tests ResNet50 training and inference, Blender rendering, Vision Transformers, LLM prompt and generation, scientific MD simulations, cryptography, and Whisper speech recognition. Results are collected per GPU core and aggregated for single‑ and multi‑GPU runs, with metrics like native_multi_gpu_result and forced_multi_gpu_sum highlighting bottlenecks.

The final BOM includes affordable X99 Intel E5‑* Xeon CPUs (e.g., E5‑2690 for $40) and a Supermicro X10DRG‑Q chassis for $200. The project demonstrates that legacy Tesla GPUs can still deliver meaningful performance for homelab AI and rendering tasks.