Benchmarking 15 'E-Waste' GPUs Against Modern AI Workloads

esologic.com / Hacker News·medium signal

A hands-on benchmark (134 points, 52 comments) runs 15 cheap secondhand/'e-waste' Tesla-class GPUs through modern inference and compute workloads to map price-per-performance for local model hosting. It quantifies where older accelerators still hold up versus where memory bandwidth and driver support break down. Practical reference for builders assembling low-cost local-inference rigs instead of paying for cloud GPUs.