Google's AI Hardware Push
Google Unveils Frozen V2 Chip and Icefish TPU Plans for Enhanced Gemini AI Efficiency
Anonymous
Key Points
1
Icefish discussions with Samsung and TSMC are non-binding for now, with no formal announcements publicly made.
2
Under Icefish, Samsung would produce the interconnect module on its 2-nanometer process to link the compute die (from TSMC) to memory, signaling a split-manufacturing approach.
3
Frozen V2 is described as embedding Gemini’s neural architecture directly into the hardware, signaling a shift from general-purpose accelerators toward model-specific hardware.
4
The Frozen V2 strategy sits within a broader industry trend, with other major players pursuing bespoke AI accelerators such as Amazon Trainium, Meta inference accelerators, and Microsoft Maia.
Google is building a new AI chip called Frozen V2 that would hardwire its Gemini model directly into the silicon, promising efficiency gains of 6 to 10 times over current chips, according to The Information via Yahoo Finance. The chip could be deployed as early as 2028 and is designed to ease a growing crunch in AI computing capacity.
Separately, Google is also working on a next-generation chip called Icefish, which would use a split-manufacturing approach — splitting production between TSMC and Samsung — on a cutting-edge 2-nanometer process, The Edge Malaysia reported.
Gemini's Brain Baked Directly Into Hardware
Most AI chips today are general-purpose accelerators. They run many different models. Frozen V2 takes a different approach. It would embed Gemini's neural architecture — the core structure of the model — directly into the chip's circuits, according to Trading Key. That means the hardware is built specifically for one job: running Gemini fast and efficiently.
The big benefit is less data movement. Moving data between chip components wastes time and power. By baking Gemini's structure into the silicon, Frozen V2 cuts that waste. The result, Google hopes, is up to 10 times more processing per unit of power compared to current TPUs, Quiver Quant reported.
Frozen V2 Would Run Alongside, Not Replace, TPUs
Google's existing chips are called TPUs, or Tensor Processing Units. The company has built them in-house since 2016. Frozen V2 would not replace TPUs. Instead, it would run separately, handling Gemini workloads specifically, according to The Edge Malaysia. Think of it as a specialist working next to a generalist.
This split strategy lets Google keep its flexible TPU line while also squeezing maximum performance out of Gemini. The 2028 target gives Google time to refine both tracks, 933 The Drive reported. No formal product announcement has been made yet.
Icefish Chip Would Split Production Between TSMC and Samsung
The Icefish chip is a separate project targeting Google's next TPU generation. Under the plan, TSMC would produce the main compute die and Samsung would handle the interconnect — the module that links the compute die to memory — all on a 2-nanometer process, The Edge Malaysia reported. Discussions with both companies are non-binding for now.
This split-manufacturing strategy signals Google wants to reduce reliance on any single chipmaker. It also highlights Samsung's growing role in cutting-edge 2nm production. A bigger Samsung presence in AI chip supply chains could pressure Nvidia, which currently dominates the market for AI accelerators.
Big Tech's Race to Build Custom AI Silicon
Google is not alone in this push. Amazon has its Trainium chip. Meta is building inference accelerators. Microsoft has its Maia chip. All of them are trying to reduce dependence on Nvidia's expensive GPUs by designing hardware tuned to their own AI models, according to Quiver Quant.
For Google, Frozen V2 represents the most aggressive step yet — moving beyond custom accelerators toward model-specific silicon. If the 6 to 10 times efficiency gains hold up, it could sharply cut the cost of running Gemini at scale. That matters enormously as AI infrastructure costs keep climbing across the industry.
Sentiment
Mostly PositiveNEGATIVENEUTRALPOSITIVE
Publishers
47
Articles
106
Reach
153
A
AnonymousAuthor