What Drives AI Compute Chips? GPUs, ASICs, and the Rise of Chinese Players

AI Compute Chip Overview

AI compute chips include GPUs, FPGAs, and ASICs such as VPU and TPU.

Market Share

GPU usage remains the largest; IDC forecasts that GPUs will still occupy about 80% of the AI chip market by 2025. Compared with traditional graphics GPUs, general‑purpose compute GPUs remove rendering functions, offering higher compute‑to‑power efficiency for model training and inference.

Key Performance Indicators

Performance, power consumption, and chip area (PPA) are the core evaluation metrics. Compute capability is measured in TOPS or TFLOPS, power efficiency is expressed as performance per watt, and smaller chip area improves yield and reduces cost while also influencing the compute density.

Nvidia GPGPU Characteristics

The dominant AI GPU family from Nvidia is defined by micro‑architecture, process node, number of CUDA cores, Tensor cores, clock frequency, memory capacity and bandwidth. Representative models include V100 (Volta), A100 (Ampere) and H100 (Hopper).

Chinese AI Chip Landscape

Domestic companies have launched a variety of AI chips covering cloud‑training, inference, and edge applications.

Cambricon

SiYuan 290 (cloud training) and SiYuan 370 (training‑inference) are built on 7nm process; SiYuan 370 is Cambricon’s first chiplet‑based AI chip with up to 256 TOPS INT8. The next‑generation SiYuan 590 is under development.

HaiGuang Information

DCU “Deep Compute No.1” uses 7nm technology, is compatible with CUDA‑like environments, and has been mass‑produced since 2021 for big‑data, AI and commercial computing workloads.

MuXi Integrated Circuit

MXN100 is a heterogeneous GPU (7nm) launched in August 2022 for inference; MXC500, also 7nm, targets AI training and general‑purpose computing and entered volume production in December 2022.

TianShu

Big Island GPGPU is a 7nm cloud‑training chip delivering 295 TOPS INT8.

Biren Technology

BR100 GPGPU adopts chiplet technology, offering >1000 TOPS FP16 and >2000 TOPS INT8 performance.

SuiYuan Technology

CloudSui i20 is a second‑generation inference accelerator card (12nm) whose efficiency rivals 7nm GPUs, achieving 256 TOPS.

KunLun

R200 AI chip, unveiled at the 2022 Intelligent Computing Summit, is based on the self‑developed XPU‑R architecture and integrates with Baidu Paddle for a friendly development environment.

PingtouGe

Develops two main lines: cloud‑centered AI chips (Yutian and Hanguang series) and RISC‑V processors for AIoT (Xuantie series). The Hanguang 800 (12nm) launched in 2019 delivers 820 TOPS.

Huawei Ascend Series

Ascend 310 focuses on inference, while Ascend 910B targets training with FP32/FP16 precision and performance comparable to Nvidia A100.

References

服务器基础知识全解终极版（第②版）

一文看懂国产AI芯片玩家

英伟达GPU龙头稳固，国内逐步追赶

字节跳动：万卡训练大模型集群技术