Chinese Moore Threads releases open-source GPU Compute Driver Bench on GitHub
Moore Threads has published GPU Compute Driver Bench as an open-source benchmark suite for GPU compute drivers. The project targets driver overhead in compute and memory heavy workloads, and it aims to provide repeatable comparisons across vendors and platforms. The code is available on GitHub under the Apache 2.0 license.
Moore Threads is a Beijing-based GPU developer founded in 2020 by former NVIDIA China executive Zhang Jianzhong. The company develops discrete GPUs and its MUSA software stack for graphics and compute workloads.
Moore Threads has officially released the GPU Compute Driver Bench , a tool that evaluates driver performance in a systematic and professional manner, helping developers, researchers, system architects, and system integrators to deeply analyze GPU driver performance in various compute- and memory-intensive scenarios.
— Moore Threads (translation)
The open-source release follows Moore Threads’ recent “Huagang” architecture update disclosed at its MUSA developer conference in December 2025 . This architecture adds FP4 to FP64 support, claims a 50% compute density increase, and targets scaling to 100,000 GPUs using MTLink, alongside new “Lushan” graphics and “Huashan” AI chips.
Source: Moore Threads
According to the GPU vendor, GPU Compute Driver Bench focuses on five areas where driver design can impact throughput and latency: scheduling, multi-stream behavior, memory operations, multi-GPU communication, and resource management. The repository describes more than 30 sub-tests, a scoring and baseline system, and automation scripts for running suites and calculating results. It also supports builds with MUSA or CUDA toolchains, and it includes CUDA-to-MUSA and MUSA-to-CUDA porting helpers tied to the musify workflow.
You can find the full guide how to install and build the test suite in Moore Threads post.
Source: Moore Threads (Weixin)
