What Drives the AI Chip Race? GPUs, ASICs, and China's Emerging Players

AI Compute Chip Landscape

AI compute chips are primarily divided into three categories: GPUs, FPGAs, and ASICs represented by VPU/TPU. According to IDC forecasts, GPUs will still occupy about 80% of the AI chip market by 2025. Compared with traditional graphics GPUs, general‑purpose AI GPUs remove rendering functions, achieving higher compute‑to‑power efficiency and becoming the dominant choice for model training and inference.

Classification by Deployment and Workload

AI chips can be split into cloud‑side and edge/terminal devices. Cloud deployments focus on high‑performance training and inference chips that handle large‑scale data analysis, model training, and bandwidth‑intensive inference. Edge and terminal chips are mainly inference‑oriented, responsible for data collection, environment perception, human‑machine interaction, and localized decision‑making.

Key Performance Indicators (KPIs)

The core metrics for evaluating AI chips are compute capability (TOPS or TFLOPS), power consumption (W), and silicon area (mm²). These three factors—often referred to as PPA—directly affect cost, efficiency, and scalability.

Compute (TOPS/TFLOPS): Measures how many trillion operations the chip can perform per second, with data types such as INT8 and FP32. Higher TOPS indicates faster processing and stronger performance.

Power: The power required for operation. Performance‑per‑watt (TOPS/W) is a critical efficiency metric.

Area: Smaller silicon area improves yield and reduces cost; the compute density (TOPS per mm²) is also a cost‑effectiveness indicator.

Global Leaders and Architecture Highlights

NVIDIA’s GPGPU remains the most widely adopted AI chip. Its performance is driven by micro‑architecture, process node, CUDA core count, Tensor core count, clock frequency, memory capacity, and memory bandwidth. Representative models include V100 (Volta), A100 (Ampere), and H100 (Hopper). These chips combine high compute density with advanced interconnects, making them the benchmark for AI workloads.

Chinese AI Chip Companies – A Panorama

Domestic manufacturers are rapidly closing the gap with several notable players:

Cambricon (寒武纪): Cloud‑training chips SiYuan 290 (512 TOPS INT8) and SiYuan 370 (256 TOPS INT8, first chiplet‑based AI chip). Upcoming SiYuan 590 is under development.

Horizon (海光信息): DCU “Deep Compute One” uses a 7 nm process, supports a CUDA‑compatible environment, and has been in volume production since 2021 for big‑data and AI workloads.

Muxi Integrated Circuit (沐曦集成电路): Heterogeneous GPU MXN100 (7 nm, INT8 inference) launched in Aug 2022; MXC500 (training/general‑purpose) entered tape‑out in Dec 2022 with full‑scale production planned for 2024.

TianShu ZhiXin (天数智芯): Big Island cloud GPGPU delivers 295 TOPS INT8 for training.

Birren Technology (壁仞科技): BR100 chip (chiplet architecture) achieves >1000 TOPS FP16 and >2000 TOPS INT8.

SuiYuan Technology (燧原科技): Cloud‑Sui i20 inference accelerator (12 nm) reaches 256 TOPS, comparable to 7 nm GPUs.

KunLun Chip (昆仑芯): R200, released at the 2022 Smart Computing Summit, uses a self‑developed XPU‑R architecture and integrates with Baidu’s PaddlePaddle platform.

PingTouGe (平头哥): Develops both data‑center GPUs (e.g., “HuanGuang 800”, 12 nm, 820 TOPS) and RISC‑V based AIoT processors.

Performance Highlights and Market Outlook

According to industry statistics, Cambricon’s SiYuan 290 leads the domestic market with 512 TOPS INT8, while Horizon’s DCU competes with NVIDIA’s A100 in certain benchmarks. Non‑listed firms such as TianShu ZhiXin and Muxi also demonstrate competitive INT8 performance (295 TOPS and 256 TOPS respectively). Major internet giants—Tencent (through SuiYuan), Baidu (KunLun), and Alibaba (through PingTouGe)—have entered the AI chip arena, further intensifying competition.

Overall, the AI chip sector remains dominated by NVIDIA globally, but a growing ecosystem of Chinese companies is expanding the supply chain, offering diverse architectures, process technologies, and performance‑per‑watt trade‑offs that could reshape the market in the coming years.