Loongson Releases First Software Version for Self-Developed GPGPU Acceleration Platform

Loongson Zhongke officially launched the first software stack for its Loongson Accelerated Computing Platform, supporting its self-developed LG200 GPGPU architecture featured in chips like the 2K3000 and 9A1000. The platform delivers OpenCL 3.0 and CUDA API compatibility alongside LacInfer, a specialized AI inference engine. Providing a full software stack from low-level drivers to AI deployment helps domestic hardware ecosystems reduce migration costs for existing CUDA and PyTorch workflows. This marks an important step toward compute chip independence and deployment in applications like computer vision and embodied AI. The release includes assembly-optimized BLAS and DNN operator libraries, a specialized GPGPU compiler, and integration with ONNX Runtime as an execution backend to run trained models without code rewrites. It also features hardware-level debugging tools to identify memory access violations, calculation anomalies, and synchronization deadlocks.