There's a peculiar subculture of hobbyists, homelab operators, and budget-conscious ML researchers who trawl eBay for decommissioned datacenter GPUs — the Tesla K80s, M40s, P40s, and P100s that once powered enterprise clusters and now sell for the price of a decent dinner. This post appears to be a rigorous benchmarking exercise across 15 of these cards, measured against workloads that actually matter in 2026: LLM inference, image generation, and modern CUDA compute.

Why does this matter? A few reasons:

• Economics of local inference are shifting. As frontier models get smaller-but-smarter (distilled 7B–30B parameter models routinely rivaling last year's flagships), the question of what hardware is "good enough" for self-hosted inference becomes genuinely interesting. A $150 Tesla P40 with 24GB of VRAM can run quantized Llama-scale models that would have required an A100 two years ago.
• The compute-per-dollar frontier isn't linear. Older architectures lose CUDA feature support (no FP8, no BF16 on some Pascal cards, no Flash Attention 2 on Kepler), which means raw FLOPS numbers mislead. Real benchmarks against real workloads expose which cards are actually usable versus which are landfill-in-waiting.
• Environmental angle. "E-waste" is doing real work in the title. Enterprises retire GPUs on tax schedules, not on capability curves. A well-documented second life extends the useful lifespan of silicon that took enormous energy to fabricate.

A technical audience would find value in the concrete numbers — tokens/second on llama.cpp, iterations/second on Stable Diffusion, power draw under sustained load, cooling requirements (many of these are passively-cooled server cards requiring 3D-printed shrouds and blower fans). The gap between marketing spec sheets and observed performance on modern software stacks is usually where the interesting story lives.

Fifteen cards is also enough to plot a curve rather than anecdotes. You start to see clusters: Kepler-era cards effectively excluded by dropped CUDA support, Maxwell hanging on for specific workloads, Pascal as the surprising sweet spot, Volta and Turing offering diminishing returns per dollar.

This is exactly the kind of practical, empirical, deeply-in-the-weeds writeup that HN historically rewards — and yet it sits at 2 points with no discussion. Probably a timing issue more than a quality one.