Loongson Technology has launched the first software version of its Loongson Accelerated Computing Platform, designed to support its proprietary general-purpose GPUs. The platform provides a full software stack, including drivers, programming environments, and AI inference tools for chips based on the LG200 series core, such as the 2K3000 and 9A1000.
The platform supports OpenCL 3.0 and CUDA programming interfaces, allowing developers to utilize existing AI models. It integrates high-performance operator libraries for BLAS and DNN, which are optimized for FP32 and INT8 data types. The platform also includes the LacInfer inference engine, which supports ONNX Runtime to enable direct deployment of models from frameworks like PyTorch and TensorFlow.
To assist with development, the platform features debugging and performance analysis tools that allow for breakpoint setting, register access, and hardware state monitoring. These tools help identify issues such as memory access violations and synchronization deadlocks. The compiler is specifically optimized for Loongson's GPU instruction set architecture to reduce migration costs.
Loongson has already optimized computer vision models, including object detection and image segmentation, for the 9A1000 chip. The platform is compatible with the entire Loongson computing chip series and Linux systems. Future support is planned for the 9A2000 and 9A3000 chips to ensure software compatibility across different generations of hardware.
The company is currently working with early customers on embodied AI device development using the 2K3000 and 9A1000 chips. Loongson aims to expand its AI software ecosystem through collaboration with industry partners and developers.
