Overview of the AI Chip Market: Architectures, Companies, and Performance Comparisons

The multi‑billion‑dollar AI chip market is scorching hot in 2023. At present, AI chips are classified by technical architecture into GPGPU, FPGA, ASIC (including VPU, TPU) and compute‑in‑memory chips.

According to their position in the network, AI chips can be divided into cloud AI chips and edge and terminal AI chips .

Cloud mainly deploys high‑compute AI training and inference chips, handling tasks such as intelligent data analysis and model training.

Edge and terminal mainly deploy inference chips, responsible for inference tasks that require independent data collection, environment perception, human‑machine interaction, and partial decision‑control functions.

According to practical goals, chips are further divided into training chips and inference chips .

Currently, 1.0‑era players such as Cambricon and Pingtouge have become listed companies offering high‑quality AI compute chips; 2.0‑era non‑listed companies like Biren Technology, Denglin Technology and Tianshu ZhiXin continue to push products; 3.0‑era startups such as Qianxin Technology and Yizhu Technology are seeking breakthroughs with compute‑in‑memory architectures.

Most AI‑chip companies now focus on edge‑side and mid‑center small‑compute scenarios such as smart security, smart city and smart medical; Biren, Pingtouge and Yizhu can also cover larger‑compute edge/center scenarios, with Yizhu boldly trying compute‑in‑memory for large‑compute workloads.

Therefore, based on architecture and application scenario classification, the following mid‑tier AI compute‑chip vendor panorama is presented:

In recent years, domestic manufacturers have made continuous breakthroughs in training‑chip hardware performance, yet a gap remains compared with market‑leading Nvidia A100.

Taking the Suoyuan YunSui T20 product as an example, its 32‑bit single‑precision floating‑point performance reaches 32 TFLOPS, surpassing the Nvidia A100’s 19.5 TFLOPS and offering power advantages, but its memory bandwidth is only one‑third of the A100, still lagging for machine‑learning and deep‑learning bandwidth demands.

AI inference chips, domestic players may catch up

Currently, domestic manufacturers such as Cambricon, Suoyuan and Kunlun have products that can directly compete with the market‑mainstream Nvidia Tesla T4; their energy‑efficiency ratio is 1.71 TOPS/W versus T4’s 1.86 TOPS/W, a relatively small gap.

With the surge of ChatGPT, the AI‑chip industry is entering the 3.0 era; architectures better suited for large models will emerge, and system‑level innovations will become the future trend, allowing early‑betting vendors to reap the benefits brought by ChatGPT.

Source: https://mp.weixin.qq.com/s/Ioq-g3TRBJdHVwN83X3h0w

Disclaimer: Thanks to the original author for the hard work. This article is re‑posted with attribution; please contact us if there are copyright issues.

Recommended Reading: For more architecture‑related technical summaries, refer to the "Architect’s Full‑Store Technical Materials Pack" (41 books) available via the "Read Original" link.

Promotional offer: The full‑store package includes server fundamentals, storage fundamentals, and other materials, priced at ¥249 (originally ¥439). Scan the QR code or follow the public account to get the free "Architect’s Full‑Store Technical Materials Pack".