Implementation:Tensorflow Tfjs MathBackendCPU

Knowledge Sources	Tensorflow_Tfjs TensorFlow.js API
Domains	Deep_Learning, CPU_Backend
Last Updated	2026-02-10 06:00 GMT

Overview

MathBackendCPU is the pure-JavaScript CPU backend for TensorFlow.js. It extends KernelBackend and provides a reference implementation that runs tensor operations entirely on the CPU using typed arrays. This backend is the fallback when no accelerated backend (WebGL, WebGPU, WASM, or Node) is available. It manages tensor data storage, reference counting, memory disposal, and supports complex number tensors.

Code Reference

Source Location

tfjs-backend-cpu/src/backend_cpu.ts (GitHub)

Key Imports

import {backend_util, BackendTimingInfo, buffer, DataStorage, DataType, engine, env,
        kernel_impls, KernelBackend, Rank, ShapeMap, Tensor, Tensor2D,
        TensorBuffer, TensorInfo, TypedArray, util} from '@tensorflow/tfjs-core';
import {assertNotComplex} from './cpu_util';

Class Signature

export class MathBackendCPU extends KernelBackend {
  public blockSize = 48;
  data: DataStorage<TensorData<DataType>>;
  constructor();
}

Exported Interface: TensorData

export interface TensorData<D extends DataType> {
  values?: backend_util.BackendValues;
  dtype: D;
  complexTensorInfos?: {real: TensorInfo, imag: TensorInfo};
  refCount: number;
}

Key Methods

write

Writes values into a new data bucket and returns a DataId. On first use in a Node.js environment, emits a warning recommending the native Node backend.

override write(values: backend_util.BackendValues, shape: number[], dtype: DataType): DataId

makeTensorInfo

Creates a TensorInfo object, handling string encoding when the dtype is 'string'.

makeTensorInfo(shape: number[], dtype: DataType,
               values?: backend_util.BackendValues|string[]): TensorInfo

read / readSync

Reads tensor data asynchronously or synchronously. For complex64 tensors, merges the real and imaginary float arrays.

override async read(dataId: DataId): Promise<backend_util.BackendValues>
override readSync(dataId: DataId): backend_util.BackendValues

disposeData

Decrements reference count and frees memory when it reaches zero. Recursively disposes complex tensor components.

override disposeData(dataId: DataId, force = false): boolean

memory

Returns memory info, noting that reported values are unreliable upper bounds due to JavaScript garbage collection.

floatPrecision

Returns 32, indicating 32-bit float precision for this backend.

I/O Contract

Method	Input	Output
`write`	`BackendValues`, shape array, dtype	`DataId` handle
`makeTensorInfo`	shape, dtype, optional values	`TensorInfo` object
`read` / `readSync`	`DataId`	`BackendValues` (typed array or Uint8Array[] for strings)
`disposeData`	`DataId`, optional force flag	`boolean` (true if memory released)
`where`	condition `Tensor`	`Tensor2D` of indices
`floatPrecision`	(none)	`32`

Usage Example

import {MathBackendCPU} from '@tensorflow/tfjs-backend-cpu';
import * as tf from '@tensorflow/tfjs-core';

// Register the CPU backend
tf.registerBackend('cpu', () => new MathBackendCPU());
await tf.setBackend('cpu');

const t = tf.tensor([1, 2, 3, 4]);
console.log(t.dataSync()); // Float32Array [1, 2, 3, 4]

Related Pages

Tensorflow_Tfjs_Layer_Utils - Utilities for printing model summaries
Tensorflow_Tfjs_Training_Utils - Training weight utilities used during model fitting

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment