GPU Power and Performance Metrics for Multiple NVIDIA Architectures

We created benchmark applications from the i) Rodinia [14] suite, ii) CUDA SDK3. Additionally, we developed a set of linear algebra-based benchmarks such as relu, softmax, scalar multiplication, scalar addition, matrix determinant, matrix transpose, and a few others. Using our data collection tool shown in GPUDevice, we executed these benchmarks on four NVIDIA GPU architectures: Tesla K80 (Kepler), Tesla V100 (Volta), RTX A4000 (Ampere), and RTX 4060 (Ada Lovelace) with various launch configurations and measured the actual power consumption. We used the NVIDIA Management Library (NVML). NVML, a C-based interface, is built for monitoring and managing various hardware states within NVIDIA Quadro and Tesla GPUs. NVML utilizes on-chip sensors for power measurement. he dataset in the GPU Power Prediction repository comprises measurements of GPU power consumption during the execution of various CUDA kernels. These measurements are essential for training machine learning models aimed at predicting GPU power usage based on kernel characteristics. Data Format The dataset is structured in CSV format, with each row representing a single kernel execution. The columns include: Kernel ID: Unique identifier for each kernel. Kernel Name: Name of the CUDA kernel. Execution Time (ms): Time taken for kernel execution in milliseconds. Power Consumption (W): Measured GPU power consumption during kernel execution in watts. Memory Usage (MB): Amount of GPU memory used during kernel execution. Thread Count: Number of threads launched by the kernel. Block Count: Number of blocks launched by the kernel. Shared Memory Usage (KB): Amount of shared memory used during kernel execution. Usage This dataset is utilized to train machine learning models that can predict GPU power consumption based on kernel characteristics. By analyzing the relationships between kernel parameters and power usage, these models can assist in optimizing GPU resource utilization and energy efficiency.