This article is compiled by Semiconductor Industry Review (ID:ICVIEWS)
Loongson’s firstGPGPUchip9A1000development is nearly complete, and tape-out will be delivered within the third quarter.
Recently, an investor asked Loongson Zhongke on an interactive platform if the9A1000 has successfully taped out.
OnSeptember15, Loongson Zhongke responded that the development of Loongson’s firstGPGPUchip9A1000 is nearly complete, and tape-out will be delivered within the third quarter. Success will depend on the test results after the tape-out returns.
It is reported that theLoongson9A1000graphics card is primarily positioned in the entry-level market andAI inference acceleration field, with performance targets aimed atAMD‘sRX 550graphics card.
Speaking of theRX 500series, this is a milestone product fromAMD, especially theRX 580 model, which is still popular among many gamers; even theRX 550, which is relatively budget-friendly, maintains a high level of activity.
Looking back at theRX 550, this graphics card usesGF14process technology, equipped withGCN 4.0architecture, containing512stream processors, matching128-bitbandwidth with2/4GB GDDR5memory configuration, usingPCIe 3.0 x8interface specification,TDPof only50W, and its overall performance is roughly comparable toNVIDIA‘sGTX 650 Ti Boost, making it a classic example of high performance and low power consumption.
According to official information, a major highlight of the Loongson9A1000graphics card chip is its support forPCIe 4.0system bus, and it is compatible with128-bit LPDDR4Xhigh-speed video memory. Although specific data on the number of computing cores, memory size, operating frequency, and power consumption have not been disclosed, the exposed structural diagram reveals eight computing clusters, supplemented by on-chip interconnect networks and secondary cache mechanisms, indicating a solid internal architecture design.
In terms of software compatibility, the Loongson9A1000fully embracesOpenGL 4.0,OpenCL 3.0, and other mainstream graphics and computingAPIs, with built-in professional video processing components, hardware decoding capabilities coveringH.264,H.265codec standards, and supports various display output protocols such asHDMI 2.1,DisplayPort 1.4, and classicVGA.
In-depth analysis of its performance specifications shows that the Loongson9A1000has a pixel fill rate of up to16GP/s (16 billion pixels per second), a texture fill rate of32GT/s (32 billion texture elements per second), and floating-point computing capabilities reaching1TFLOPS (1 trillion operations per second) atFP32 precision,64GFLOPs (640 billion operations per second) atFP64 precision, and in theINT8 integer computing environment, it soars to32TFLOPS (32 trillion operations per second). These metrics collectively outline a powerful computing platform that balances graphics rendering and deep learning acceleration.
In the past,GPU, as its full nameGraphics Processing Unit, was primarily used for tasks such as graphics rendering. Today,GPU has been assigned a more “versatile” role. For example, in the current boomingAI large model, when developers train models, they need to process trillions of data. If we compare the process of handling this massive data to the “plowing” of the digital age,GPU is like simultaneously starting hundreds or thousands of high-performance automated tractors to process these “fields” in a very short time, making the operation of theAI model more efficient.
Loongson has been conducting preliminary research onGPU since 2016, with the original intention of providing supporting capabilities forCPU. At that time, theGPU industry was not as hot as it is now, and during the promotion of LoongsonCPU, issues related toGPU kept emerging. For example, the supply channels for importedGPU chips were unstable, and embeddingGPU in desktop scenarios was not feasible, etc. These factors significantly impacted the functionality, performance, and cost-effectiveness of Loongson computers. Thus, Loongson concluded a rule:Any company that makesCPU must have its ownGPU.
Initially, the Loongson team started almost from scratch in theGPU field, but with the belief that “it cannot be harder thanCPU” they began to explore. However, in-depth research revealed thatGPU, as an acceleration system for graphical applications, involves a lot of application-layer related background knowledge and lacks clear documentation likeCPU, making it extremely difficult to learn. Therefore, the team started with research on graphical algorithms, going through simulator architecture design research, simulator validation, logic design, and functional validation, taking5 years to launch the first generation of graphicalGPU architecture. Subsequently, they spent another2 years on two minor iterations, upgrading to Loongson7A2000 and Loongson2K2000 and pushing them to market.
After the first generationGPU entered the product iteration stage, Loongson quickly initiated the upgrade work for the second generationGPU architecture, aiming to pushGPU towards the fourth stage of development, fromgraphics processor (GPU to general-purpose graphics processor (GPGPU).
9A1000 is not the only graphics card project from Loongson. The company is also developing9A2000, which is aimed at mid-to-high-end graphics cards for desktop and server applications, optimizing and enhancing its capabilities. TheGPU core is upgraded to the third generation architecture, with further improvements in computing power per unit area, graphicsAPI support for OpenGL4.6, virtualization support added, tensor units supporting more data types,GPU scale x4 (vs 9A1000), single-precision floating-point computing power of5Tflops,INT8 AI computing power of160TOPS, memory bandwidth of256GB/s, supporting dual-chip interconnection, overall performance doubling, reaching internationally advanced levels for the same process generation. Additionally, Loongson also plans to launch9A3000 as a follow-up product to9A2000, but currently, there is no specification information available.
*Disclaimer: This article is the original author’s creation. The content of the article reflects their personal views, and our reprint is only for sharing and discussion, not representing our endorsement or agreement. If there are any objections, please contact us.
