Google is reportedly working on developing its own chip that will have part of the architecture of its Gemini AI model etched into it. Internally, it is called Frozen v2 and, according to information leaked so far, it should run several times more efficiently while consuming less power. What else do we know?
Frozen v2 as an answer to problems
News of the new Frozen v2 chip was first reported by The Information, which cited people directly involved in the project. According to Tom's Hardware and TechCrunch, which picked up the story, the new chip should generate six to ten times as many tokens per unit of energy consumed as Google's current chips, known as TPUs. It is scheduled to launch in 2028.
And why now? Google's urgency is no coincidence. Tom's Hardware reports that it may also be a response to the current shortage of computing power for AI. The shortage was reportedly so severe that Google Cloud even had to turn down orders from external customers for a while.
Gemini frozen in silicon
Conventional TPU or GPU chips can run any model loaded onto them. However, while running it, they must make a whole series of decisions and process huge volumes of data. This results in high energy consumption.
Frozen v2, however, bypasses this. The part of Gemini's architecture that controls these decisions is directly “frozen” into its transistors. As a result, the chip performs fewer steps and transfers less data when processing each query. This could significantly reduce response times and energy consumption. Perhaps enough to pave the way for entirely new applications.
From specific data to general architecture
The original version of Frozen, whose development was led by Google DeepMind chief scientist Jeff Dean according to Tom's Hardware, was supposed to go even further and etch the data of a specific Gemini model directly into the chip. However, Google ultimately backed away from this and decided to fit only its general architecture into the chip. The reason is that although silicon tied to a specific version of the model would be even faster and more efficient, it would have a very short lifespan.
Frozen v2 therefore “freezes” only the architecture, essentially the model's skeleton. This allows the chip to remain usable across multiple versions of Gemini. However, even that only applies until Google changes the underlying architecture.
Google remains silent, investors are pleased
Google has yet to officially confirm or deny the project. According to TechCrunch, it merely stated that the company is constantly exploring new innovations. But the market has responded in its own way. Following the report, Alphabet shares rose by about 3% on Monday. Google plans to spend $180 billion to $190 billion on AI development this year and needs to demonstrate that the investment will pay off.
Google is not alone
The idea of permanently “etching” a model into a chip is not new. Tom's Hardware points to Canadian startup Taalas, which unveiled the HC1 chip with a permanently embedded Llama 3.1 8B model back in February.
Other AI companies are also pursuing their own approaches. According to TechCrunch, OpenAI unveiled its first custom Jalapeño chip in June, while Anthropic is in talks with Samsung about chip production. They are all driven by the same motivation: to operate more efficiently and gain independence from Nvidia's chips, which currently dominate the entire market.
Source: Tom's Hardware, TechCrunch
