Hasty Briefsbeta

Bilingual

Benchmarking 15 "E-Waste" GPUs with Modern Workloads

10 hours ago
  • #GPU Benchmarking
  • #AI Performance
  • #Homelab Hardware
  • 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.