Deep-Dive: Google Tensor G6 NPU Architecture & Gemini Nano Integration

Underneath the **Google Tensor G6**, the NPU sub-system transitions to an 8-core decoupled matrix tile architecture. Each core contains dedicated INT4 and FP16 vector execution pipelines tuned specifically for transformer attention operations.

Key Technical Developments

To overcome mobile DRAM bandwidth limits during LLM token generation, Tensor G6 incorporates a 32MB System Level Cache (SLC) paired with LPDDR5X memory running at 9.6 Gbps, allowing **Gemini Nano 2** to process up to 45 tokens per second locally.

Industry Impact & Outlook

Dynamic power gating ensures that background agentic tasks execute within a strict 1.2W envelope, balancing AI functionality with device thermal management.