Web AI Compute Benchmark

This repository contains interesting AI workloads running on the Web, using WebAssembly (Wasm), WebGPU, and in future WebNN or other underlying technologies. Those workloads help us to evaluate the performance of the implementation, e.g., in browsers or JavaScript and WebAssembly runtimes.

The runner, which allows to select workloads, including ones from external sources, and collects and displays metrics, is based on the Speedometer runner. See the Speedometer repo for a more detailed explanation, e.g., in which phases workloads are run and measured.

Setup Instructions, How to Run Workloads

• Prerequisites: NPM, node. npm install to install the dependencies of the runner.
• Building the individual workloads: Run npm run build in the root directory. This will install dependencies and build all workloads in resources/. Alternatively, you can also manually run npm install and npm run build inside each resources/<workload-group>/ sub-directory to produce output in resources/<workload-group>/dist/.
• Starting the web server: npm run dev in the root directory.
• Most important files:
  • Workloads are in resources/transformers-js/ and resources/litert-js.
  • Shared files are in resources/shared/, which is depended-upon as a local package.
  • The default suite / tests to run are in resources/default-tests.mjs.

How to Run Individual Workload

• If you have not done that yet, run npm install and npm run build inside resources/<workload-group>/ to produce output in dist/.
• npm run dev in the root directory to start the server.
• Navigate to http://localhost:8080/resources/<workload-group>/dist/<workload>.html (e.g. http://localhost:8080/resources/transformers-js/dist/feature-extraction-cpu.html) then do manualRun() in the developer console.

Developer Mode and Custom Parameters

The runner supports a number of options to control how the benchmarks are run. They can be set via URL parameters or more conveniently via the developer menu.

To enable the developer menu, append ?developerMode to the URL, e.g. http://localhost:8080/?developerMode. Then click on the red box in the top left corner to show this:

You can select workloads by tag (e.g. all #wasm or #webgpu workloads) and change parameters of the runner in it. Changing them in the developer menu should be immediately reflected in the URL, producing e.g. http://localhost:8080/?iterationCount=1&tags=wasm

A full list of the runner options can be found in params.mjs.

Benchmark Methodology

• Iterations: By default, the benchmark suite runs for 5 iterations (controlled by iterationCount URL parameter or Iterations in the developer menu).
• Inferences per Iteration: Each iteration runs the inference 5 times by default (controlled by subIterationCount URL parameter or Sub-iterations in the developer menu).
• Consistent Input: The same input data is used for all inferences across all iterations. The input data and model are loaded once during initialization to ensure measurements focus on compute performance.
• Initialization Time: Time spent on model loading and data preparation is not included in the benchmark execution time. Initialization happens once upfront before measurement begins.

Interpreting Results

The benchmark UI displays results at three levels of detail:

• Workload Overview (Top of Page): Displays the total time per workload. This is calculated as the sum of all sub-iteration times, averaged across all iterations.
• Sub-Iteration Averages (Scrolling Down): Displays the average time for each specific sub-iteration slot across all iterations (e.g., the average of sub-iter-1 from Iteration 0 to Iteration 4).
• Detailed Tables (Expanding "Table"): Displays the raw inference time for every single sub-iteration in every iteration in a grid format.

Each individual sub-iteration represents a single inference run (a single call to the model's execution function using the same input data).

Contributing

See the CONTRIBUTING file for how to help out.

Source Code Headers

Every file containing source code must include copyright and license information. This includes any JS/CSS files that you might be serving out to browsers. (This is to help well-intentioned people avoid accidental copying that doesn't comply with the license.)

BSD 2-clause header:

Copyright 2025 Google LLC

Use of this source code is governed by a BSD-style
license that can be found in the LICENSE file or at
https://developers.google.com/open-source/licenses/bsd

Workloads from third party projects or files from the original Speedometer runner may have different, but compatible licenses (e.g., Apache 2). The respective subdirectories should have the appropriate LICENSE file.