NVIDIA Tesla e-waste GPUs: V100 beats T40, M60 kills Whisper for $50
$75 P100s and $50 M60s still crush modern AI workloads—benchmarked with real results.
I spent a year building GPU coolers and a custom benchmarking tool to test decommissioned NVIDIA Tesla GPUs (K80, M10, M40, M60, P40, P100, V100, T40) on LLMs, vision, Whisper, and Blender. The V100 (16GB, under $200) surprised me—its performance hangs right up there with the much more expensive T40. For LLMs, the P40 beats the P100. The $50 M60 is a Whisper beast, beating even the V100. With cheap X99 Xeon motherboards, stacking cards generally scales linearly, but mixing generations can let slower cards bottleneck faster ones in LLM setups. Software limits are easy to work around (compile older tools from source), and you can just power off the box when not doing AI tasks—making e-waste a killer homelab option.
- V100 (16GB, ~$200) performs on par with the T40 in modern AI workloads, making it the best value e-waste GPU.
- P40 beats P100 for LLMs; M60 ($50) outperforms all others for Whisper audio transcription.
- Linear performance scaling when stacking GPUs; cheap X99 Xeon setups work well with no major bottlenecks.
Why It Matters
Professionals can build high-VRAM AI homelabs for under $500 using e-waste GPUs, democratizing access to machine learning workloads.