Google is reportedly developing a dedicated server chip for Gemini in an attempt to support the demand for AI reasoning for sustained growth at a lower cost. According to the report, the chip codenamed Frozen v2 could be deployed as early as 2028.

No more just TPU Upgrade

According to the report, Frozen v2 is not the normal iterative of Google's existing TPU. Unlike a generic AI chip that can run multiple models, this product will directly anchor part of Gemini's architecture into hardware.

As a result, chips can reduce double counting when processing requests and the cost of back-to-back transmission of data between memory. Engineers expect to produce 6 to 10 times more efficient token generation of the existing TPU than their unit power.

Calculator pressure drives self-study.

Google advanced this project in the context of the fast-growing demand for AI services. It was mentioned that Google had previously experienced a problem of inadequate server capacity to meet even the computing needs of some external clients.

The user-side experience may not change immediately if the dedicated chip is successfully landed, but Gemini ' s running costs may decrease significantly. This will have a direct impact on the competition between Google and OpenAI, Anthropic and the Chinese AI Laboratory.

The factory continues to emerge from its dependence on Young Waida.

This also reflects the fact that large AI companies are continuing to reduce their dependence on the GPU. Generic GPU remains the core hardware for current AI training and reasoning, but for super-large model service providers, dedicated chips have a better chance of reducing long-term costs.

  • Chip codename is Frozen v2
  • The earliest deployment time points to 2028
  • Target efficiency increased to 6 to 10 times

In addition to Google, Meta, Amazon, Microsoft and OpenAI are also promoting self-study chips. According to the report, Frozen v2 is still in the exploratory phase, critical designs have not yet been finalized and Google has not formally confirmed the existence of the project.

Additional information:It was also reported that the chip would not be open to external cloud customers because its hardware design was customized around Gemini and it was difficult to run other models directly.