Google was revealed to be "writing" Gemini into the chip? Inference energy efficiency is improved by up to 10 times

Compiled by Zhidongzhi | Editor by Eggplant | Cheng Qian Zhidongzhi reported on July 21 that yesterday, The Information quoted people familiar with the matter as reporting that Google is developing a new AI server chip Frozen v2 for the Gemini model. The chip plans to directly integrate part of the Gemini model architecture into the chip to improve AI reasoning efficiency by reducing calculation decisions and data handling during operation. According to people familiar with the matter, Google engineers estimate that based on the number of tokens that can be output per unit of power consumption, the efficiency of Frozen v2 may be 6 to 10 times that of Google's latest self-developed AI chip. The chip is informally called "Frozen v2" within Google. People familiar with the matter said that Google hopes to alleviate the serious shortage of AI computing power. The shortage of computing power has caused conflicts within Google and forced Google Cloud to reject orders from some external customers. People familiar with the matter also revealed that Google may deploy this chip as soon as 2028, but the company is still determining the main functions of Frozen v2, how the various parts work together, and how much model information will be solidified in the chip. The Information also mentioned that Frozen v2 is not the first time Google has explored model-specific chips. Previously, Jeff Dean, chief scientist of Google DeepMind, had planned to solidify the model weights directly into the chip, but because the chip of this solution could only adapt to a specific version of the model and the life cycle was too short, it was ultimately shelved. 1. Write some of the Gemini logic into the chip to reduce chip data handling. Most current mainstream AI chips emphasize versatility and programmability. For example, NVIDIA GPUs can support a variety of AI models and computing tasks. This type of chip needs to support a variety of different AI models, so a large amount of calculation scheduling and decision-making needs to be completed according to the model requirements during operation. Google TPU is similar. Although it is optimized for AI, it also needs to take into account the running needs of different models. Frozen v2 uses different