Benchmarking 15 decommissioned NVIDIA enterprise GPUs with modern AI workloads reveals surprising performance and value, despite being EOL'd and less power efficient. The V100, in particular, outperforms the more expensive T40 in several tests. These findings have significant implications for homelab enthusiasts and datacenter operators looking to repurpose 'e-waste' GPUs for modern workloads. However, it is essential to consider the potential drawbacks, such as power efficiency and compatibility issues, when repurposing these GPUs.
By understanding the performance and limitations of these GPUs, individuals can make informed decisions about their use in various applications, including AI, computer vision, and scientific computing. [1] The benchmarking process involved testing the GPUs with a custom Dockerized suite, which included tests such as ResNet50 training and inference, Blender GPU rendering, and llama.cpp large language model processing. [2] The results show that the V100, despite being an older and less expensive GPU, performs similarly to the T40 in many tests. [3] Additionally, the findings suggest that the power efficiency of these GPUs may not be a significant concern for homelab enthusiasts, as the cost savings can outweigh the increased power consumption.
Overall, the benchmarking of these 'e-waste' GPUs provides valuable insights into their potential for modern workloads and highlights the importance of considering the technical and business implications of repurposing decommissioned hardware.
