Addressing AI Computing Power Shortage
Deployment Expected as Early as 2028

Google is reportedly developing a customized semiconductor optimized for its artificial intelligence (AI) model, Gemini. The company appears to be accelerating in-house chip development to address the shortage of computing power required for AI processing.


Reuters Yonhap News

Reuters Yonhap News

View original image

On July 20 (local time), The Information cited sources who revealed that Google is working on a new server chip known as "Frozen v2," which could be deployed as early as 2028. This semiconductor is being developed separately from Google's Tensor Processing Unit (TPU). According to The Information, it is designed to reduce data movement and respond to user queries more quickly.


Unlike the existing TPU, which is versatile and can handle various AI models much like Nvidia's graphics processing units (GPUs), "Frozen v2" is being designed to incorporate some of Gemini's data processing methods and model structures directly into the chip. It reportedly applies an idea proposed by Jeff Dean, Google's Chief Scientist.


Google expects that running Gemini on Frozen v2 could increase the number of tokens processed per unit of power by six to ten times compared to current capabilities, while also improving response times. However, sources noted that Google is currently reviewing how much information to embed in the chip in advance to balance flexibility for multiple AI tasks with the efficiency of a Gemini-specific chip.


It has been reported that Google's pursuit of Frozen v2 is motivated by a serious internal shortage of computational resources. This issue has led to conflicts among internal departments, and Google Cloud has had to decline some requests from external customers.


The production scale of Frozen v2 is expected to be significantly smaller than that of existing TPUs. Currently, the project is thought to be at a stage focused on verifying technical feasibility, rather than full commercialization. This is because, even within Gemini, changes in underlying architecture between different versions could make it challenging to utilize the chip.



A Google spokesperson explained, "We are continually researching and experimenting with new technologies to deliver maximum performance and efficiency for our users and customers," adding, "By designing hardware and software together from the ground up, we are building an integrated system optimized for real workloads."


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing