The rapid deployment of large language models (LLM) is fueling the overall AI trend. After witnessing the growth of LLM applications such as ChatGPT, Taiwan-based IC design houses are no longer skeptical of AI demand potential and are devoting themselves to the development of related technologies.
Taiwan-based IC design houses such as MediaTek, Novatek Microelectronics, Realtek Semiconductor, Himax Technologies, Elan Microelectronics, Sunplus Technology, and Egis Technology have already invested a significant amount into AI and the development of related products.
Manufacturers have been focusing on image-processing products, according to industry players. MediaTek is using its AI processing unit (APU) engine to supplement computing for gaming image quality and various camera upgrades.
Novatek, Realtek, and Himax have also developed AI-related products for image quality and surveillance image modules. Elan and Egis have AI computing chips aimed at automotive image functions. Sunplus has developed its own AI image computing platform.
Sources pointed out that companies are focusing on image processing because there is already clear end demand for applying AI to imaging. Compared with other applications that may be bound to specific application products, AI imaging can be used in almost any application.
The success of LLMs is a departure from the previous development path of AI personal assistants by major US-based companies. In addition to ChatGPT, it has been reported that more field-specific small GPT models will emerge in the future.
Bloomberg recently launched BloombergGPT, an AI model trained on the financial market.
As more small-scale models are launched, the more they can be directly imported to edge devices without the need for cloud computing. This will bring about more diversified application development for mainstream devices such as mobile phones and notebooks. It also means that all devices in the future will need an AI chip to realize these applications.
The key to AI becoming a more integrated part of users' everyday lives lies in whether it can be edgified. If computing can be done directly on the edge, it would open up a huge business opportunity for the chip industry.
The industry still needs time to trim down the relevant models to a scale suitable for edge device computing, according to sources. However, some companies have already been successful on the chip side.
Qualcomm has already implemented offline computing functionality for Stable Diffusion on its flagship Snapdragon 8 Gen 2 mobile phone chip.
In addition to operating through traditional computing methods such as CPUs, GPUs, and even APUs, Taiwan-based IC designers are also interested in processing related models via ASICs, which are more flexible and can be made more compact.
Sources believe more industry players will launch corresponding ASIC products for AI functions such as LLMs.
