BREAKING
Google's 'Frozen' Chip to Cut Gemini Cost
0
x
low estimate
0
x
high estimate
0
possible deployment
How 'Frozen' Works
1
Freeze architecture in silicon
↓
2
Reduce data movement
↓
3
Keep weights updatable
↓
4
Speed up inference
Frozen v2 vs Existing TPUs
Frozen v2
new line
●
Model-specific ASIC
●
Experimental scale
●
Targets Gemini inference
TPUs
since 2016
●
General-purpose accelerator
●
8th-gen lineup April 2026
●
Training and inference
Split Reaction, Unconfirmed Plan
Projections, Not Measured Results
AI NEWS BLITZ
A report says Google is building a chip that hardwires Gemini into silicon to slash inference costs.