Benchmarking 15 “E-Waste” GPUs with Modern Workloads
The article 'Benchmarking 15 “E-Waste” GPUs with Modern Workloads' tests a range of older Tesla GPUs, including models like the Tesla P40, K80, and M40, against modern benchmarks. The workloads cover AI inference (using frameworks like TensorFlow and PyTorch), 3D rendering (Blender), and general compute (Geekbench). The Tesla P40, based on the Pascal architecture with 24GB of VRAM, performs particularly well in AI inference tasks, often matching or exceeding newer consumer cards in throughput for batch processing. The K80, despite its age, shows utility for double-precision compute. The author notes that these GPUs can be purchased for under $200 on the secondary market, making them an attractive option for hobbyists or small-scale deployments. However, they lack modern features like hardware ray tracing and have higher power consumption. The benchmarks reveal that for certain non-gaming workloads, these 'e-waste' GPUs remain viable, challenging the notion that older hardware is obsolete.
Developers can repurpose cheap, older Tesla GPUs for AI inference, reducing hardware costs for experimentation.