Google Tensor Chips — DarWINN NPU Access

Direct-access tooling for the Google Tensor NPU (DarWINN), reverse-engineered across two generations: Tensor G2 (Pixel 7, codename Janeiro) and Tensor G3 (Pixel 8, codename Rio).

Why this exists

I was trying to get LLM inference running on the phone and wanted to use the Tensor NPU to speed it up. To do that I had to reverse-engineer how the NPU is actually driven, since it has no public API — only Google Camera and signed system services are allowed to use it.

Along the way I learned the NPU doesn't help with LLMs, on either generation: language-model decoding is memory-bandwidth-bound, and the NPU shares the same memory bus as the CPU, so it ends up no faster (~1.81 ms/block NPU vs ~1.57 ms CPU on G2; ~2.01 ms vs ~1.66 ms on G3, for a real transformer FFN block). For my own use case that was a dead end on both chips — the LLM runs better on the CPU.

But the access tooling itself works, and the NPU is a strong accelerator for the workloads it was built for (vision / CNN models, where the weights fit on-chip). It didn't serve my project, but others doing on-device vision or NPU research might find it useful — so I'm releasing it under MIT.

What's here

If you just want a local LLM running on your phone, start with on-device-llm/ — it's the useful, no-root part. The g2-pixel7//g3-pixel8/ folders are for people curious about why the NPU isn't used and how the locked hardware was accessed anyway.

The two chips' NPUs expose the same DarWINN API v2 surface (AddInput/AddOutput/Submit) but differ in one important way: G2 hands back a raw tensor pointer, while G3 wraps tensors in an opaque Buffer* handle that must be accessed through its own SizeBytes/MapToHost/FlushCache API. See g3-pixel8/README_DEMO.md for what that changes in practice.

What I found, in short

• The NPU can be driven directly from a rooted device, at the same privilege as the Camera app — on both G2 and G3.
• Custom models run on it (demonstrated on G2): int8 weights, uint8 I/O, per-tensor quantization → loadable via the google-edgetpu NNAPI accelerator (per-channel weights get rejected).
• It's a compute accelerator, not a memory one. Great for vision/CNN (weights fit the on-chip SRAM). No help for LLM decode (bandwidth-bound, shares the LPDDR5 bus with the CPU) — confirmed on both generations.

Environment

on-device-llm/ needs nothing but Termux — no root. The NPU folders need a Magisk-rooted Pixel 7 (Tensor G2) or Pixel 8 (Tensor G3) · Android 14/16 · Frida 17.x · target process com.google.android.GoogleCamera · libedgetpu_util.so — device-owner access to your own hardware, for research and learning.

License

MIT. Shared as-is, no warranty, in case it helps someone.