Can Chinese GPUs Close the Gap with NVIDIA? 2023 GPGPU Landscape Analysis

The article originates from the “GPGPU Research Framework and Compute Analysis (2023)” and identifies high‑precision floating‑point computation and the CUDA ecosystem as the core barriers for GPGPU development.

From a raw compute perspective, domestic GPU products still lag one generation behind foreign counterparts, and the software/ecosystem gap with NVIDIA’s CUDA is even more pronounced.

Specific product comparisons include:

BR100 from Wall‑Rim Technology surpasses NVIDIA A100 in FP32 single‑precision performance but does not support FP64 double‑precision.

TianGai100 from TianShu ZhiXin exceeds A100 in FP32 but falls short in INT8 integer performance.

DCU from HaiGuang achieves FP64 double‑precision at roughly 60% of A100’s performance, comparable to its level four years ago.

Beyond hardware, software ecosystems are crucial. NVIDIA’s CUDA commands a 90% share of the global GPU market, while most Chinese firms rely on the open‑source OpenCL, which requires extensive time to build a robust ecosystem. By contrast, AMD spent nearly a decade (since 2013) developing its ROCm platform before gaining noticeable traction.

Short‑term implications of the U.S. ban on high‑end GPUs suggest potential setbacks for China’s AI computing, supercomputing, and cloud industries. In the meantime, alternatives include non‑banned NVIDIA/AMD products and domestically produced mid‑ to high‑performance CPUs, GPUs, and ASICs.

Long‑term, the vast domestic market and the “Xinchuang” (information technology innovation) demand are expected to accelerate the localization of AI chips, significantly increasing the domestic share of AI chip production and enabling gradual convergence with international standards.

Recommendations for domestic manufacturers focus on independent innovation and building a self‑sufficient ecosystem. Key areas and representative companies are:

Chips: Loongson (PC CPUs), HaiGuang (servers, DCU), Jingjia Micro (graphics GPUs), Cambricon (ASICs), Lattice (server memory interfaces).

PCBs: Shenghong Technology, Xingsen Technology, Huadian Co.

Advanced packaging: Tongfu Microelectronics, Yongsi Electronics, Changdian Technology, Changchuan Technology.

Overseas leaders: NVIDIA, AMD, Intel, Micron.