Loongson Technology is actively engaged in the research and production of GPUs and related products. It has launched several chips integrated with self-developed GPU cores and is accelerating the deployment of dedicated GPGPU chips.
1. Iteration and Application of Self-Developed GPU Cores
Loongson Technology insists on independently developing GPU core technology and has launched multiple generations of self-developed GPU cores. In the early stages, the first-generation graphics GPU architecture, the LG100 series, completed development in the first half of 2022 and was integrated into the 7A2000 bridge chip, supporting graphics rendering and basic computing. Subsequently, the second-generation GPU core, LG200, further optimized graphics algorithms and performance, integrated into chips like the 2K3000, and possesses certain AI inference capabilities, which have been introduced into customer scenarios.
2. Progress of the First Dedicated GPGPU Chip 9A1000
In June 2025, Loongson Technology officially disclosed its first dedicated GPGPU chip, the 9A1000, positioned as a low-cost solution for the graphics card/AI accelerator card market. The development of this chip is nearly complete, with tape-out scheduled for the third quarter (testing to be completed within 2025). In terms of performance, the 9A1000’s graphics performance is comparable to the AMD RX550 (supporting OpenGL 4.0), with AI computing power of INT8 greater than 32 TOPS. Compared to the LG200 core integrated in the 2K3000, performance has improved by more than five times, and AI computing power reaches 40 TOPS. Additionally, through core upgrades (doubling the graphics pipeline, increasing the main frequency by 25%), reducing the area of stream processors by 20%, and lowering light-load power consumption by 70%, improvements in computing density and energy efficiency have been achieved.
3. Mid- to High-End Product Planning
In addition to the 9A1000, Loongson Technology has planned mid- to high-end GPGPU products:
- 9A2000
: Targeting the mid- to high-end graphics card/server market, it will integrate four 9A1000 chips, upgrade the graphics API to OpenGL 4.6, support virtualization and floating-point data types such as BF16 and FP8, with single-precision floating-point computing power of 5 TFLOPS, INT8 AI computing power of 160 TOPS, memory bandwidth of 256 GB/s, and support for dual-chip interconnection (doubling overall performance), aiming to reach internationally advanced levels under the same process, expected to be launched in 2027.
- 9A3000
: A higher-performance GPGPU, planned to adopt advanced Xnm technology, with performance improvements of 3-5 times over the 9A2000, further enhancing Loongson’s high-end layout in the GPGPU field.
4. Ecosystem and Application Synergy
Loongson Technology’s GPU products create a synergistic effect with CPUs, collaborating with partners to launch hardware bases such as training servers, AI training cards, and AI inference cards, supporting mainstream frameworks like TensorFlow and PyTorch, and adapting to operating systems such as Loongnix, Tongxin, and Kirin. They are widely used in AI scenarios such as smart cities, smart transportation, and computer vision, building a self-sufficient AI ecosystem from edge to cloud.
