useTPU

Source: packages/sdk/src/ai/useTPU.ts

The Tensor TPU is reachable through AICore (Gemini Nano via ML Kit Prompt API in @pixelkit-labs/mlkit) or LiteRT. This hook reports what is verifiably installed; real on-device inference metrics (latency, token counts, TTFT) live in useGeminiNano(). benchmarkTPU() runs a real 256×256 JS matmul and reports it as CPU fallback, clearly labelled.

Verified on Pixel 11 Pro: AICore 0.release.prod_aicore_20260723.00_RC11, Private Compute Services 1.0.release.962568596. Requires the <queries> declaration in the module manifest (Android 11+ package visibility). Package and feature detection runs once on mount.

Signature

function useTPU(): {
 aicoreInstalled: boolean;
 aicoreVersion: string | null;
 privateComputeServicesVersion: string | null;
 hasNpuFeature: boolean | null;
 activeDelegate: 'Tensor TPU' | 'NPU' | 'GPU' | 'CPU Fallback';
 isHardwareAccelerated: false;
 lastInferenceLatencyMs: number | null;
 throughputTokensPerSec: number | null;
 memoryFootprintMB: number | null;
 cpuFallbackLatencyMs: number | null;
 isBenchmarking: boolean;
 benchmarkTPU: () => Promise<TPUAcceleration>;
 source: TelemetrySource;
};

Outputs

Field	Type	Description
aicoreInstalled	boolean	Whether com.google.android.aicore is installed. false when absent, or when package visibility is not declared.
aicoreVersion	string | null	AICore versionName, the host process for Gemini Nano. null when not installed.
privateComputeServicesVersion	string | null	Version of com.google.android.as.oss, which brokers private model downloads. null when not installed.
hasNpuFeature	boolean | null	The android.hardware.neural_processing_unit feature flag as the device declares it. null before the read.
activeDelegate	'Tensor TPU' | 'NPU' | 'GPU' | 'CPU Fallback'	Which executor the last measurement used. Becomes 'CPU Fallback' once benchmarkTPU() has run, because that is what it actually measures.
isHardwareAccelerated	false	Always false: this hook runs nothing on the TPU. Accelerated inference is useGeminiNano.
lastInferenceLatencyMs	number | null	Always null here. See useGeminiNano().lastLatencyMs for a measured on-device figure.
throughputTokensPerSec	number | null	Always null here. See useGeminiNano().lastDecodeTokensPerSec.
memoryFootprintMB	number | null	Always null: AICore does not report model memory to apps.
cpuFallbackLatencyMs	number | null	Duration in ms of the last benchmarkTPU() matmul, explicitly labelled a CPU number. null until run.
isBenchmarking	boolean	true while the matmul occupies the JS thread.
source	TelemetrySource	'hardware' when the native module is present, 'unavailable' otherwise.

Functions

Function	Inputs	Returns	Description
benchmarkTPU()	none	Promise<TPUAcceleration> — { activeDelegate: 'CPU Fallback', isHardwareAccelerated: false, lastInferenceLatencyMs: number, throughputTokensPerSec: null, memoryFootprintMB: null }	Runs a real 256×256 float32 matrix multiply on the JS thread and returns its duration. Labelled CPU fallback because no part of it touches the TPU.