- Decommissioned NVIDIA Tesla GPUs (e.g., K80, P100, V100) offer affordable VRAM for homelab use, despite being end-of-life and less power-efficient.
- A benchmarking suite was developed to evaluate these GPUs across various AI, rendering, and compute workloads, using Docker for compatibility.
- Results show that V100 performance competes with newer T40, P40 is recommended for LLMs, and M60 excels at audio transcription with Whisper.
- Multi-GPU scaling tests indicate linear performance gains in many workloads without significant overhead, beneficial for dense GPU nodes.
- CPU choice impacts performance slightly, with single-core speed favoring some tasks, while motherboard differences had minimal effect.
- The author plans to build a GPU node with V100s, an E5-1680 CPU, and Asus X99-WS motherboard for projects like streamarize.