Recently, an investor asked Loongson Technology on an interactive platform whether the 9A1000 has successfully taped out.On September 15, Loongson Technology responded that the development of its first GPGPU chip, the 9A1000, is basically complete, and tape-out will be delivered within the third quarter. The success of the tape-out will depend on the test results after it returns.
It is reported that the Loongson 9A1000 graphics card is primarily positioned for the entry-level market and the AI inference acceleration field, with performance targets aimed at AMD’s RX 550 graphics card.
Speaking of the RX 500 series, this is a milestone product from AMD, especially the RX 580 model, which is still popular among many gamers; even the relatively affordable RX 550 version maintains a high level of activity.
Looking back at the RX 550, this graphics card uses a GF14 nanometer process technology, equipped with GCN 4.0 architecture, containing 512 stream processors, matched with a 128-bit bandwidth of 2/4GB GDDR5 memory configuration, and adopts the PCIe 3.0 x8 interface specification. Its TDP is only 50W, and its overall performance is roughly comparable to NVIDIA’s GTX 650 Ti Boost, making it a classic example of high performance and low power consumption.
According to official information, a major highlight of the Loongson 9A1000 graphics chip is its support for the PCIe 4.0 system bus, and it is compatible with 128-bit LPDDR4X high-speed video memory. Although specific data on the number of compute cores, memory size, operating frequency, and power consumption has not yet been disclosed, the exposed structural diagram reveals that it has laid out eight major compute clusters, supplemented by an on-chip interconnect network and a secondary cache mechanism, indicating a solid internal architecture design.
On the software compatibility front, the Loongson 9A1000 fully embraces industry-standard graphics and computing APIs such as OpenGL 4.0 and OpenCL 3.0, and includes professional-grade video processing components, with hardware decoding capabilities covering H.264 and H.265 codec standards. It also supports various display output protocols, including HDMI 2.1, DisplayPort 1.4, and the classic VGA.
In terms of performance specifications, the Loongson 9A1000 boasts a pixel fill rate of up to 16GP/s (16 billion pixels per second), a texture fill rate of 32GT/s (32 billion texture elements per second), and floating-point computing capabilities reaching 1TFLOPS (1 trillion operations per second) at FP32 precision, and 64GFLOPs (64 billion operations per second) at FP64 precision. In an INT8 integer computing environment, it can soar to 32TFLOPS (32 trillion operations per second). These metrics collectively outline a powerful computing platform that balances graphics rendering and deep learning acceleration.
In the past, GPUs, as their full name Graphics Processing Units suggests, were primarily used for graphics rendering tasks. Today, GPUs have been assigned a more versatile role. For example, in the context of the currently popular AI large models, when developers train models, they need to process trillions of data points. If we compare the process of handling this massive data to “tilling the land” in the digital age, GPUs are like hundreds or thousands of high-performance automated tractors working simultaneously to process these “plots” in a very short time, making the operation of AI models more efficient.
Loongson began its GPU research and development work in 2016, initially aimed at providing support for CPUs. At that time, the GPU industry was not as hot as it is now, and during the promotion of Loongson CPUs, various issues related to GPUs emerged. For example, the supply channels for imported GPU chips were unstable, and embedded GPUs were not usable in desktop scenarios. These factors significantly affected the functionality, performance, and cost-effectiveness of Loongson computers. Consequently, Loongson concluded that any company that produces CPUs must have its own GPU.
Initially, the Loongson team started almost from scratch in the GPU field, but they began exploring with the belief that “it cannot be harder than CPU development.” However, in-depth research revealed that GPUs, as acceleration systems for graphics applications, involve a lot of application-layer related background knowledge and lack clear documentation like CPUs, making the entry-level learning curve extremely steep. Therefore, the team started with research on graphics algorithms, going through simulator architecture design, simulator verification, logic design, and functional verification, taking five years to launch the first generation of graphics GPU architecture. Subsequently, they spent another two years on two minor iterations, upgrading to the Loongson 7A2000 and Loongson 2K2000 and bringing them to market.
After the first generation of GPUs entered the product iteration stage, Loongson quickly initiated the upgrade work for the second generation GPU architecture, aiming to push the GPU into its fourth developmental stage, transitioning from a graphics processor (GPU) to a general-purpose graphics processor (GPGPU).
The 9A1000 is not the only graphics card project from Loongson. The company is also developing the 9A2000, which is aimed at mid-to-high-end graphics cards for desktop and server applications, focusing on optimization and material stacking to strengthen its capabilities. The GPU core will be upgraded to a third-generation architecture, with further improvements in computing power per unit area, graphics API support for OpenGL 4.6, virtualization support, tensor units supporting more data types, and a GPU scale of x4 (vs 9A1000), with single-precision floating-point computing power of 5Tflops, INT8 AI computing power of 160TOPS, memory bandwidth of 256GB/s, and support for dual-chip interconnection, effectively doubling overall performance to reach internationally advanced levels for the same process generation. Additionally, Loongson plans to launch the 9A3000 as a follow-up product to the 9A2000, but currently, there is no specification information available.
PS: Source material from the internet, for reference only. If you find it helpful, don’t forget to like and share by clicking“❤”Little Red Heart
Recommended Reading
1. National Procurement Center 2024 Domestic Computer Bulk Centralized Procurement Operating System Data Released
2. National Procurement Center 2024 Domestic CPU Computer Procurement Data Released
3. Domestic Xinchuang Peripheral Adaptation Issues and Adaptation List
4. In the context of the cancellation of the Xinchuang catalog, how to determine the autonomy and controllability of Xinchuang products
5. Interpretation of the Xinchuang Standard “Government Procurement Demand Standards” article
6. A Brief Discussion on Xinchuang Adaptation Verification and Xinchuang Laboratory Construction
7. First Batch of Security and Reliability Evaluation Results (1-2023)
8. Can domestic Xinchuang operating system computers run Windows programs?9. Xinchuang = Localization? An article to quickly understand Xinchuang and localization10. Xinchuang and Open Source Conflict? Can open-source software be used in Xinchuang projects? (Including open-source software list)11. Interpretation of the Security and Reliability Evaluation Results (2024 No. 1) document12.CentOS countdown to shutdown in less than a month: “Is the domestic operating system ready?” Arrangements!13. Latest list of security and reliability evaluation results for domestic Xinchuang CPU and operating system (1-2024)14. Summary of domestic Xinchuang PCs15. How to choose Xinchuang computers