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

AI Compute Chip Types

AI compute chips are classified into three categories: general‑purpose GPUs, reconfigurable FPGAs, and ASICs such as VPU/TPU. GPUs dominate the market; IDC forecasts roughly 80% share in 2025.

Key Performance Indicators (KPIs)

Evaluation metrics focus on compute performance (TOPS/TFLOPS), power consumption, and silicon area (PPA). Compute is measured in trillion operations per second for INT8 and FP32. Performance‑per‑watt assesses efficiency; smaller die area reduces cost and improves yield.

Compute (TOPS/TFLOPS): Number of trillion operations a chip can perform per second, typically reported for INT8 and FP32 precision.

Power: Wattage required for operation; performance‑per‑watt is a key efficiency indicator.

Area: Chip die size, influencing cost and manufacturing yield.

Global Market Landscape

NVIDIA’s GPGPU is the most widely adopted AI chip worldwide. Performance drivers include micro‑architecture, process node, CUDA core count, Tensor core count, clock frequency, memory capacity and bandwidth. Notable architectures are Volta (V100), Ampere (A100) and Hopper (H100).

Chinese AI Chip Companies Overview

Cambricon (寒武纪): SiYuan 290 (512 TOPS INT8) and SiYuan 370 (256 TOPS INT8, chiplet). SiYuan 590 is under development.

HaiGuang Information (海光信息): DCU “Deep Compute One” built on 7 nm, compatible with a CUDA‑like ecosystem, in volume production since 2021.

MuXi Integrated Circuit (沐曦集成电路): MXN100 (7 nm, INT8 inference) and MXC500 (training, 2022).

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

BiRui Technology (壁仞科技): BR100 chiplet architecture, >1000 TOPS FP16 and >2000 TOPS INT8.

SuiYuan Technology (燧原科技): CloudSui i20 inference accelerator (12 nm) with 256 TOPS, comparable to 7 nm GPUs.

KunLun Chip (昆仑芯): R200 AI chip (2022) based on self‑developed XPU‑R architecture, integrated with Baidu PaddlePaddle.

PingTouGe (平头哥): HanGuang 800 data‑center GPU (12 nm) delivering 820 TOPS, plus RISC‑V AIoT processors.

Huawei’s Ascend series (e.g., Ascend 310 for inference, Ascend 910B for training) provide FP32/FP16 performance comparable to NVIDIA A100.

Performance Highlights

Cambricon’s SiYuan 290 reaches 512 TOPS INT8, while SiYuan 370 achieves 256 TOPS INT8. HaiGuang’s Deep Compute One matches several parameters of NVIDIA A100. TianShu ZhiXin’s BI chip delivers 295 TOPS INT8. MuXi’s MXN100 was taped out in August 2022.