Recently, the domestic processor manufacturer Loongson Technology responded on an interactive platform that its first GPGPU dedicated chip — the 9A1000 has successfully completed design and development, and is expected to tape out within this quarter.
It is reported that the Loongson 9A1000 will adopt an advanced chip design architecture, featuring high throughput parallel computing capabilities, aimed at meeting the demands of AI inference, edge computing, and certain high-performance computing scenarios. The 9A1000 is positioned as an entry-level graphics card, supporting AI acceleration, with performance comparable to the AMD RX550, possessing certain inference capabilities, and supporting OpenGL 4.0; in terms of terminal AI computing performance, its INT8 performance exceeds 32 TOPS. Additionally, the 9A1000 will integrate a video processing module, supporting H.264/H.265 encoding and decoding.
According to the company’s chief engineer and head of general GPU processor development, Su Menghao, the Loongson 9A1000 has achieved a performance leap of over five times compared to its predecessor — the second-generation GPGPU core LG200 integrated in the 2K3000 chip, with its AI computing power reaching 40 TOPS (40 trillion operations per second).
This chip benefits from a comprehensive upgrade of the 9A1000’s GPU core, with a doubled graphics pipeline, a 25% increase in clock frequency, and a 20% reduction in the area of stream processors, leading to a 70% decrease in light load power consumption. Most importantly, the overall scale of the 9A1000’s GPU has reached four times that of the previous LG200 core. This chip is expected to be launched in 2026.
The company further disclosed its GPGPU product line planning blueprint — the Loongson 9A2000, an independent graphics card chip aimed at the mid-to-high-end market. This chip will integrate four mature 9A1000 chips within an advanced package, upgrade the graphics API to OpenGL 4.6, and add support for virtualization, in addition to supporting INT8 and INT4, the tensor units will also increase support for BF16, FP8, and other floating-point data types.
In terms of computing power, the 9A2000 achieves 5 TFLOPS of single-precision floating-point performance, 160 TOPS of INT8 AI computing power, and a memory bandwidth of 256 GB/s. One of the most notable features of the 9A2000 is its support for dual-chip interconnection, which can further double the overall performance, reaching an internationally advanced level under the same process.
Loongson Technology’s chairman, Hu Weiwu, has emphasized, “Achieving lower costs, higher performance, and a better ecosystem through self-reliance.” He revealed that the subsequent 9A3000 will be a higher-performance GPGPU, and if advanced Xnm technology is adopted, performance could increase by 3-5 times.
For a long time, the central processing unit (CPU) and graphics processing unit (GPU) have constituted the two core components of modern computing systems. Loongson Technology has achieved significant accomplishments in the general computing field with its self-developed LoongArch instruction set and the “Loongson” series CPUs, widely applied in critical industries such as government affairs, power, and transportation. The successful development of the Loongson 9A1000 will further promote the localization of GPUs.
The completion of this research and development and the planned tape-out in the third quarter indicates that the chip design has entered the final verification and manufacturing stage. After successful tape-out, Loongson will conduct sample testing and performance validation, paving the way for subsequent mass production and commercial applications.
THE END
Follow “Electronic Engineering Magazine” and add the editor’s WeChat
Regional groups are now open, please send messages 【Shenzhen】【Shanghai】【Beijing】【Chengdu】【Xi’an】 to the public account
