GPU Performance Prediction Challenge

🎯 Goal

Predict the Graphics Performance Index (GPI) - a composite metric representing the overall performance of a graphics card - based on its hardware specifications. Your model will learn from a dataset of over 12,000 GPUs, covering a wide range of architectures, memory configurations, and clock speeds.

This is a regression problem. The more accurately you can estimate GPI from specs like core count, memory bandwidth, and clock frequencies, the better your score.

📁 Dataset Description

You are provided with 17 numerical features describing the hardware of each GPU:

• cores: Number of shading units / CUDA cores
• tensor_cores: Number of Tensor cores (for AI acceleration)
• rt_cores: Number of Ray Tracing cores
• sm_count: Number of Streaming Multiprocessors
• base_clock: Base clock frequency (MHz)
• boost_clock: Boost clock frequency (MHz)
• memory_clock: Memory clock frequency (MHz)
• ram: Video memory size (GB)
• bus_width: Memory bus width (bits)
• fp16_performance: Half-precision floating-point performance (TFLOPS)
• fp32_performance: Single-precision performance (TFLOPS)
• fp64_performance: Double-precision performance (TFLOPS)
• bf16_performance: BFloat16 performance (TFLOPS)
• tf32_performance: TF32 performance (TFLOPS)
• int8_performance: INT8 inference performance (TOPS)
• multi_gpu: Multi-GPU support flag (0/1)
• api_cuda: CUDA compute capability version

Target variable: gpi_value - a continuous value representing the Graphics Performance Index (higher is better).

The dataset is derived from real-world GPU specifications and benchmark results, cleaned and merged for this competition.

📊 Data Splits

• train.csv: 10,000 rows containing 18 columns (id + 17 features + gpi_value). This is the training data with ground truth.
• test.csv: 2,709 rows containing 17 columns (id + 17 features). This is the test data without the target.
• sample_submission.csv: 2,709 rows containing 2 columns (id + gpi_value filled with mean). This serves as a template for your predictions.

📈 Evaluation Metric

Submissions are evaluated using Root Mean Squared Error (RMSE). Lower RMSE means better performance. The final score on the leaderboard is based on the Private test set.

📨 Submission Format

For each id in the test set, you must predict a single numeric value for gpi_value. Your submission file must be a CSV with exactly two columns: id and gpi_value.

File format requirements: - Header: id,gpi_value (exactly as shown) - Delimiter: comma (,) - Encoding: UTF-8 - No extra columns – only these two are allowed.

Your predicted gpi_value should be a floating-point number (any precision is acceptable). The order of rows does not matter, but every id must appear exactly once.