Google is reportedly developing a new AI processor codenamed Frozen v2, designed specifically for its Gemini series of large‑language models. The Information says the chip could deliver six to ten times the performance per watt of the company’s current silicon, a claim that sent Alphabet’s shares up about 1.5% when the report surfaced.
Why a dedicated chip matters for Gemini
Off‑the‑shelf graphics processors often carry components that AI workloads don’t use. A startup that buys a standard GPU may find its rendering cores idle while the inference cores handle model execution, leading to wasted power and higher costs. Custom silicon can eliminate those unused blocks, or replace them with circuits tuned for the target workload.
Google already offers its own custom AI accelerators, the Tensor Processing Units (TPUs), through its cloud platform. The latest TPU variants—TPU 8t for training and TPUi for inference—are built to handle a broad range of models. According to the report, Frozen v2 will go a step further by being tailored to the architecture of Gemini models.
One of the efficiency routes cited is a reduction in the number of calculations required to run Gemini. By cutting down on arithmetic steps, the chip could also support operator fusion, a technique that merges multiple operations into a single, faster computation.
Technical goals and rollout timeline
Data movement is another focus. When a neural network exceeds the on‑chip memory of a GPU, parts of the model must be swapped between off‑chip storage and the processor, slowing execution. Frozen v2 is expected to carry enough memory to keep Gemini fully on‑chip, removing the need for such transfers.
The design may also integrate with Google’s existing TPU cluster infrastructure. Current TPU 8i systems use optical circuit switches to route data among devices. Keeping compatibility would allow Google to deploy Frozen v2 without a major overhaul of its data‑center fabric.
Related: Alibaba unveils new AI model Qwen3.8
Google aims to begin shipping the chip to its data centers by 2028, giving the company time to refine the silicon and align it with the evolving Gemini roadmap.
Looking ahead, the move reflects a broader industry trend toward application‑specific hardware. If the performance claims hold, Frozen v2 could set a new benchmark for energy‑efficient AI inference, especially for models that dominate Google’s own services.
Beyond raw efficiency, a Gemini‑optimized silicon layer can incorporate specialized instruction sets that map directly onto the model’s attention mechanisms, reducing the overhead of generic compute pipelines. This alignment can lead to tighter latency budgets, a key factor for real‑time applications such as search personalization or conversational agents that rely on rapid response times.
The prospect of embedding sufficient on‑chip memory also opens the door to new training paradigms. With the entire model resident in fast local storage, iterative weight updates can proceed without the penalty of frequent off‑chip fetches, potentially shortening the time required to fine‑tune Gemini on domain‑specific data.
Maintaining coherence with the existing TPU ecosystem means that existing software stacks, including the widely used JAX and TensorFlow libraries, can be extended rather than rewritten. Developers would benefit from a seamless migration path, preserving investment in tooling while gaining the performance uplift offered by the bespoke hardware.
Finally, the strategic timing of a 2028 rollout aligns with Google’s broader roadmap for generative AI services. By the time the chip reaches production, Gemini is expected to have matured through multiple model iterations, ensuring that the hardware’s capabilities are matched to the most advanced versions of the architecture.
