Benchmarking 15 decommissioned NVIDIA GPUs for modern workloads

An independent researcher benchmarked 15 decommissioned NVIDIA enterprise GPUs, including the K80, M40, P100, V100, and T40, to evaluate their viability for modern homelab AI tasks. Using a custom Dockerized suite across LLMs, computer vision, Blender, and Whisper, the study challenges the notion that these EOL cards are obsolete.

• The 16GB V100 emerged as the sweet spot, matching the performance of the more expensive T40.
• For LLM inference using Pascal architecture, the P40 outperformed the P100.
• The M60 proved highly capable for audio transcription with Whisper, beating the V100 and costing only around $50.
• Performance scaling is linear when stacking cards in a 4U chassis, though mixing generations can bottleneck faster GPUs.
• Any cheap X99 motherboard paired with a high-lane Xeon CPU provides sufficient bandwidth for these GPUs.

The findings suggest that decommissioned enterprise hardware remains a cost-effective source of VRAM for homelab enthusiasts willing to compile older software from source.