Chinese GPU maker Moore Threads has launched an open-source GPU Compute Driver Bench on GitHub, inviting developers to contribute to its performance evaluation suite. The tool, released under the Apache 2.0 license, supports both Moore Threads' MUSA drivers and CUDA-compatible GPU drivers, enabling fair cross-platform comparisons.

This comprehensive benchmarking suite focuses on realistic workloads across diverse compute and memory scenarios aligned with real-world usage patterns. The toolkit employs a multi-dimensional evaluation system that weighs driver performance, resource management, and execution efficiency. It provides standardized metrics and baselines for methodical hardware and software optimization comparisons.

The automated scoring system enables granular subsystem testing alongside holistic end-to-end performance assessment, with regression tracking across driver and hardware versions. This release follows Moore Threads' January launch of TileLang-MUSA, which promised to reduce code volume by approximately 90%. The company's emphasis on CUDA compatibility isn't surprising given its founder Zhang Jianzhong's background as a former NVIDIA China executive.