BREAKING
Google's 'Frozen v2' Etches Gemini in Silicon
0
x
tokens/watt low
0
x
tokens/watt high
0
target year
How Freezing the Model Works
1
Freeze architecture
↓
2
Etch weights in silicon
↓
3
Cut data movement
↓
4
Boost efficiency
TPU vs Frozen v2
TPU
since 2016
●
Programmable
●
Training and inference
Frozen v2
~2028
●
Fixed-function ASIC
●
Inference only
The Bet and the Risk
Google Unconfirmed, Details Thin
AI NEWS BLITZ
Google is reportedly building a chip that hardwires its Gemini models straight into silicon.