AMD, Intel, and Nvidia are now selling top-of-the-line computing GPUs to Chinese customers without permission from the US government and Chinese developers of GPUs for artificial intelligence (AI) and high-performance computing (HPC) you can’t. Applications do not have access to state-of-the-art production capabilities, and smaller GPU designers reportedly have opportunities to take advantage of various types of generative AI. Digi Times.
The rise of China’s chip design sector in recent years has shaped two vectors of domestic GPU development. Datacenter GPUs designed to address the AI and HPC megatrends, and traditional GPUs designed primarily for client PCs but capable of some datacenter-specific workloads. The latter seems more likely to succeed in the current situation.
Small GPU developers can handle big problems
Innosilicon, Jingjia Microelectronics, and Moore Threads are probably the best-known developers of Chinese gaming graphics processors. Game graphics rely on single-precision floating-point (FP32) computational throughput, so GPUs from Innosilicon, Jingjia, and Moore’s Threads support this data format. As of now, these GPUs can hardly claim a place among the best graphics cards. Additionally, to accommodate various artificial intelligence and machine learning applications, these companies have fine-tuned their hardware to support lower-precision data formats (FP16, BF16/8, INT8, INT4, etc.) and specific instruction should be supported. For matrix and vector processing. Some have already done this, some have not yet.
Jingjia has been offering gaming GPUs based on their proprietary architecture since 2014, and their latest JM9 series GPUs from 2021 promise to offer performance levels similar to Nvidia’s GeForce GTX 1080. For now, Jingjia’s GPUs can’t accommodate his AI/DL/ML applications, although the company told DigiTimes that it is working on AI-capable GPUs for various applications such as speech recognition and natural language processing. , did not provide further details.
Innosilicon will unveil its 1st generation Fenghua (Fantasy) discrete graphics processor with Imagination Technologies’ PowerVR microarchitecture in late 2021, followed by its 2nd generation Fantasy GPU for low-power applications in mid-2022. , announced the development of the 3rd generation Fantasy with ray. Last year tracking support. Innosilicon’s original Fantasy GPU supports both FP32 and INT8, and its drivers support the latest application programming interfaces for computing, including DirectX, Vulkan, OpenCL, Caffe 1.0, TensorFlow 1.1.2, and ONNX. increase.
Moore Threads hardware is probably better suited for AI. The company’s latest Chunxiao graphics processors support FP32, FP16, and INT8 precisions, at least for some AI/DL/ML workloads, assuming they also support the appropriate instruction sets can. The company also said that the MTVerse platform will enable developers to build applications such as big data, AI training and inference, speech recognition, and visual recognition.
some background
According to VMR data cited by DigiTimes, the total market for graphics processing units available in China reaches $4.739 billion, accounting for 18.7% of the global market share. It’s unclear if the $4.739 billion includes data center GPUs. The market is projected to grow to USD 34.56 billion by 2027, with a 7-year compound annual growth rate (CAGR) of 32.8%.
There are about 10 GPU developers in China. Two of them, Biren Technology and Tianshu Zhixin Semiconductor, are purely focused on processors for AI and HPC applications, and their GPUs aren’t exactly for processing graphics. However, producing these processors requires cutting-edge manufacturing technology, and TSMC must obtain an export license from the US government to produce GPUs for Biren and Tianshu Zhixin.
On the other hand, Jingjia, Innosilicon, Moore Threads, and many more or less universal GPU developers that render games, but can also serve AI and technical computing applications if you get the right hardware capabilities. .
Until recently, Chinese designers of gaming GPUs have not been inclined to build AI- and HPC-oriented processors that rival the solutions of big companies such as Nvidia, Intel, AMD, Biren and Tianshu Zhixin. With the future of China-based AI and HPC GPU developers uncertain, and the limited ability of AMD, Intel, and Nvidia to serve Chinese customers, they have reassessed their plans, You may come up with a chip that can accommodate your needs. DL/ML hardware.
