BREAKING
Gemma 4 E2B Runs on Intel NPUs
0
x
Prefill speed
0
x
Decode speed
0
x
Perf-per-watt
OpenVINO backend for LiteRT
1
LiteRT runtime
↓
2
OpenVINO backend
↓
3
Core Ultra NPU
CPU vs GPU prefill on Lunar Lake
CPU prefill
435
GPU prefill
3751
LiteRT-LM vs llama.cpp memory
LiteRT-LM
lean
●
About 1.5-2 GB
●
Less heat, fewer kills
llama.cpp
●
About 4-5 GB
●
Higher footprint
Always-on local LLM on AI PCs
AI NEWS BLITZ
Google and Intel bring Gemma 4 E2B to Intel NPUs in Core Ultra AI PCs.