Google is developing a specialized server AI chip, Frozen v2, with the goal of partially integrating the Gemini architecture directly into the hardware. This approach is expected to provide a manifold increase in energy efficiency compared to current solutions.
What Happened
Google is working on project Frozen v2—a specialized AI chip implementing a "model-in-silicon" strategy. According to available data, the deployment of the new architecture is scheduled for 2028. The primary technical objective is to achieve a 6–10x increase in energy efficiency compared to current TPUs, which would significantly increase the number of tokens processed per unit of power consumed.
Context
Currently, the industry relies on general-purpose accelerators, such as Google's TPUs or solutions from Nvidia. However, Google is striving for extreme vertical integration, where the architectural features of Gemini neural networks will be designed with the physical parameters of specific silicon in mind. This requires the company to ensure strict compatibility between future software versions of the models and the chip's hardware base.
Why It Matters for the Industry
The shift toward a specialized model-in-silicon architecture could radically reduce capital expenditures for building and operating data centers, as well as help address the shortage of computing power. This sets a new trend for deep software-hardware integration, intensifying competition between general-purpose accelerator providers and developers of closed vertical ecosystems.
Why It Matters for Users
For end users, the implementation of this project means that AI assistants will become significantly faster and cheaper to use, which is critical for tasks requiring real-time operation, such as voice interfaces. However, it is worth noting that model architecture may become less flexible due to the tight coupling with hardware.
What Is Not Yet Known / Limitations
The project is in the early stages of development, and official benchmarks or detailed technical specifications of the architecture are currently unavailable. The information is based on media reports, and actual efficiency metrics may differ from those projected.
Sources
Author
Look at AI, Editorial Team