BenchmarkXPRT Blog banner

Category: image processing

WebXPRT 5: Starting to assemble the pieces

Justin Greene
By Justin Greene
on November 6, 2025

In our last blog post, we shared the exciting news that we’re currently working on WebXPRT 5. In that post, we described some of the ways that WebXPRT may evolve with the release of WebXPRT 5. In today’s post, we’ll revisit some of the points of emphasis from the last post and focus on potential workload changes in a bit more detail.

With any benchmark development project, there are always technical challenges you need to iron out. That is especially true with a cross-platform, browser-based benchmark like WebXPRT. Because we’re in the middle of exploring the technical feasibility of a few of the options we’ll mention, we’re not yet ready to say for certain that all these features will be available in the initial WebXPRT 5 release. We can, however, now paint a clearer picture of the overall direction we’re headed.

In the section below, you’ll find updated info on where we stand with respect to some of the key development focal points we discussed in our last post. If there’s an item from that post or previous posts that we didn’t mention below—such as updating the test harness—it doesn’t mean that we’re dropping that goal. We’re just focusing on workloads today.

One of our key goals with WebXPRT 5 is providing more AI-related workloads. In past blog posts, we’ve discussed the growing importance of local, browser-side AI. With WebXPRT 5, we’re investigating two ways that we can expand WebXPRT’s AI portfolio: 1) updating existing WebXPRT 4 AI-oriented workloads, and 2) adding all-new AI workloads.

Here are some possible ways those AI-related changes may play out in both categories:

Updating existing WebXPRT 4 AI-oriented workloads

  • Splitting the existing Organize Album using AI workload’s timed tasks—face detection and image classification—into two independent workloads.
  • Updating the face detection and image classification tasks with the latest versions of the OpenCV.js computer vision and machine learning libraries.
  • Updating the Caffe deep learning framework for the face detection task.
  • Updating the ONNX-based SqueezeNet machine learning model for the image classification tasks.
  • Updating the version of the Tesseract.js OCR engine that WebXPRT uses in the Encrypt Notes and OCR Scan workload. 

Potentially adding all-new AI workloads (either core or experimental workloads)

  • We’re exploring the idea of including a workload that uses an AI-powered segmentation model to blur the background of a video call.
  • We’re exploring the feasibility of including a local LLM chat workload.
  • We would eventually like to include a WebGPU-based web AI framework for a computer vision workload.

In addition to the goal of adding more AI, we previously discussed the possibility of adding non-AI WebGPU workloads. As a web API, WebGPU enables web-based applications—such as image-based GenAI and inference workloads—to directly access the graphics rendering and computational capabilities of a system’s GPU. In the future, WebXPRT 5 could use that technology to execute complex 3D rendering workloads.

We hope today’s post gives you a better sense of where WebXPRT 5 may be headed. We want to reemphasize that while we are actively investigating the possible changes mentioned above, nothing is set in stone. As the pieces start to fall into place, we’ll provide more information here in the blog.

If you have any questions or comments about WebXPRT 5, please feel free to contact us!

Justin

Posted in AI, benchmark, Benchmark metrics, Benchmarking, Benchmarks in general, BenchmarkXPRT, browser performance, Browser-based benchmarks, Caffe, computer vision, Cross-platform benchmarks, face detection, Future of performance evaluation, image classification, image processing, JavaScript, large language models, ONNX Runtime Web, Performance benchmarking, SqueezeNet, Web AI, WebGPU, WebXPRT, WebXPRT 4 | Tagged , , , , , , , , , , , , , , , , , , , |

You asked, and we heard you: WebXPRT 5 is on the way!

Justin Greene
By Justin Greene
on October 9, 2025

We’re excited to announce that WebXPRT 5 is officially on the way! Since we launched WebXPRT 4 in February 2022, it’s proven to be an exceptionally successful and reliable go-to benchmark for OEM labs, the tech press, and individual users alike—to the tune of over 644,000 runs to date. In past blog posts, we’ve discussed new features and possible auxiliary workloads that we contemplated adding to WebXPRT 4. As we’ve considered user comments and suggestions, changes in web technology, and how we can best position WebXPRT as a relevant browser benchmark in the future, however, it became clear that it was time for an all-new WebXPRT.

Now that we’ve announced WebXPRT 5, the first question for many existing WebXPRT users may be, “When will WebXPRT 5 be available?” We’re not yet ready to share an anticipated WebXPRT 5 release date, but we can share that a lot of groundwork is already complete, and the remaining work is moving along rapidly. We’ll continue to issue updates here in the blog as we reach important milestones.

The second question for many existing WebXPRT users may be, “How will WebXPRT change?” We’re not yet ready to share extensive details about WebXPRT 5’s workloads—rest assured that we will as soon as we can firm up everything—but we can share a few key guidelines we tried to follow in our WebXPRT 5 design. Each of these points of emphasis is a result of feedback we’ve received from labs, as well as features that users have asked for.

  • Provide more AI-related workloads. In past blog posts, we’ve discussed the growing importance of local, browser-side AI. WebXPRT 4 already includes timed AI tasks in two of its workloads: the Organize Album using AI workload and the Encrypt Notes and OCR Scan workload. We’re working on ways to expand WebXPRT’s AI portfolio in the next version.
  • Add WebGPU workloads. As a web API, WebGPU enables web-based applications—such as image-based GenAI and inference workloads—to directly access the graphics rendering and computational capabilities of a system’s GPU. We hope to incorporate WebGPU measures in WebXPRT 5.
  • Improve WebXPRT’s utility as a tool for test labs, publications, and engineering analysis.
    • Update the workloads with longer operations. Many of WebXPRT’s existing workloads no longer challenge cutting-edge consumer hardware as much as many of us would like. Testing labs have asked for longer and more demanding workloads. We’re working on incorporating workloads that are accessible enough to be run by a broad range of devices yet challenging enough to allow performance differentiation among high-end systems.
    • Enable more precise performance measures. Labs and testers have also asked for more granular insight into the workloads to help with engineering-level performance analysis. Currently, some WebXPRT 4 workload scores include multiple timed tasks. If we separate those compound scores so that each workload reports results from only one timed task, users will be able to more precisely assess how well a device performs while handling specific operations. We’re looking into this approach.
  • Modernize the harness to make it more flexible and to speed future work. WebXPRT 4’s current harness works with server-side sessions on a LAMP (Linux, Apache, MySQL and PHP) stack. If we implement the harness via JavaScript on the client side, it will pave the way for faster development and testing cycles in the future.

We expect WebXPRT 5 to carry on the WebXPRT legacy of reliability and real-world relevance, while providing users with compelling new workloads and features. As has been our habit with new benchmark releases, however, we won’t force anyone to change versions anytime soon. Instead, we will continue to make WebXPRT 4 available for quite some time after WebXPRT 5 goes live.

If you have any questions or comments about WebXPRT, please let us know!

Justin

Posted in AI, Application-based benchmarks, benchmark, Benchmark metrics, Benchmarking, BenchmarkXPRT, BenchmarkXPRT development community, browser performance, Browser-based benchmarks, Cross-platform benchmarks, Future of performance evaluation, GPU, image classification, image processing, inference, LAMP, on-device AI, Performance benchmarking, Web AI, web API, Web-based testing, WebGPU, WebXPRT, WebXPRT 3, WebXPRT 4 |

Speaking of potential future WebXPRT workloads

Justin Greene
By Justin Greene
on October 31, 2024

In recent blog posts, we’ve discussed several types of potential future WebXPRT workloads—from an auxiliary AI-focused workload to a WebXPRT battery life test—and many of the factors that we would need to consider when developing those workloads. In today’s post, we’re discussing other types of workloads that we may consider for future WebXPRT versions. We’re also inviting you to send us your WebXPRT workload ideas!

Currently, the most promising web technology for future WebXPRT workloads is WebAssembly (Wasm). Wasm is a binary instruction format that works across all modern browsers, provides a sandboxed environment that operates at native speeds, and takes advantage of common hardware specs across platforms. Wasm’s capabilities offer web developers significant flexibility in running complex client applications within the browser.

We first made use of Wasm in WebXPRT 4’s Organize Album and Encrypt Notes workloads, but Wasm has the potential to support many more types of test scenarios. Here are just a few of the use-case categories that Wasm supports:

  • Gaming
  • Image and video editing
  • Video augmentation
  • CAD applications
  • Interactive learning portals
  • Language translation

Those categories and the possibilities they open for additional workloads are exciting! When thinking through possible new workload scenarios, it’s important to remember that workload proposals need to fit within a set of basic guidelines that uphold WebXPRT’s strengths as a benchmark. You can read about those guidelines in more detail in this blog post, but in short, new workloads ideally should

  • be relevant to real-life scenarios
  • have cross-platform support
  • clearly differentiate in their performance between different types of devices
  • produce consistent and easily replicated results

After testing with WebXPRT or reviewing the list of use cases that Wasm supports, have you considered a new workload or test scenario that you would like to see? If so, please let us know! Your ideas could end up playing a role in shaping the next version of WebXPRT!

Justin

Posted in battery life, benchmark, Benchmarking, BenchmarkXPRT, browser performance, Browser-based benchmarks, Collaborative benchmark development, Cross-platform benchmarks, Future of performance evaluation, image processing, Performance benchmarking, WebAssembly, WebXPRT, WebXPRT 4, What makes a good benchmark? | Tagged , , , , , , , , , |

A clearer picture of WebXPRT 4

Justin Greene
By Justin Greene
on October 7, 2021

The WebXPRT 4 development process is far enough along that we’d like to share more about changes we are likely to make and a rough target date for publishing a preview build. While some of the details below will probably change, this post should give readers a good sense of what to expect.

General changes

Some of the non-workload changes in WebXPRT 4 relate to our typical benchmark update process, and a few result directly from feedback we received from the WebXPRT tech press survey.

  • We will update the aesthetics of the WebXPRT UI to make WebXPRT 4 visually distinct from older versions. We do not anticipate significantly changing the flow of the UI.
  • We will update content in some of the workloads to reflect changes in everyday technology. For instance, we will upgrade most of the photos in the photo processing workloads to higher resolutions.
  • In response to a request from tech press survey respondents, we are considering adding a looping function to the automation scripts.
  • We are investigating the possibility of shortening the benchmark by reducing the default number of iterations from seven to five. We will only make this change if we can ensure that five iterations produce consistently low score variance.

Changes to existing workloads

  • Photo Enhancement. This workload applies three effects to two photos each (six photos total). It tests HTML5 Canvas, Canvas 2D, and JavaScript performance. The only change we are considering is adding higher-resolution photos.
  • Organize Album Using AI. This workload currently uses the ConvNetJS neural network library to complete two tasks: (1) organizing five images and (2) classifying the five images in an album. We are planning to replace ConvNetJS with WebAssembly (WASM) for both tasks and are considering upgrading the images to higher resolutions.
  • Stock Option Pricing. This workload calculates and displays graphic views of a stock portfolio using Canvas, SVG, and dygraph.js. The only change we are considering is combining it with the Sales Graphs workload (below).
  • Sales Graphs. This workload provides a web-based application displaying multiple views of sales data. Sales Graphs exercises HTML5 Canvas and SVG performance. The only change we are considering is combining it with the Stock Option Pricing workload (above).
  • Encrypt Notes and OCR Scan. This workload uses ASM.js to sync notes, extract text from a scanned receipt using optical character recognition (OCR), and add the scanned text to a spending report. We are planning to replace ASM.js with WASM for the Notes task and with WASM-based Tesseract for the OCR task.
  • Online Homework. This workload uses regex, arrays, strings, and Web Workers to review DNA and spell-check an essay. We are not planning to change this workload.

Possible new workloads

  • Natural Language Processing (NLP). We are considering the addition of an NLP workload using ONNX Runtime and/or TensorFlowJS. The workload would use Bidirectional Encoder Representations from Transformers (BERT) to answer questions about a given text. Similar use cases are becoming more prevalent in conversational bot systems, domain-specific document search tools, and various other educational applications.
  • Message Scrolling. We are considering developing a new workload that would use an Angular or React.js to scroll through hundreds of messages. We’ll share more about this possible workload as we firm up the details.

The release timeline

We hope to publish a WebXPRT 4 preview build in the second half of November, with a general release before the end of the year. If it looks as though that timeline will change significantly, we’ll provide an update here in the blog as soon as possible.

We’re very grateful for all the input we received during the WebXPRT 4 planning process. If you have any questions about the details we’ve shared above, please feel free to ask!

Justin

Posted in benchmark, Benchmark metrics, Benchmarking, Browser-based benchmarks, Collaborative benchmark development, Cross-platform benchmarks, Future of performance evaluation, HTML5, image classification, image processing, Performance benchmarking, Web-based testing, WebAssembly, WebXPRT, webxprt 4 | Tagged , , , , |

Considering WebAssembly for WebXPRT 4

Justin Greene
By Justin Greene
on March 25, 2021

Earlier this month, we discussed a few of our ideas for possible changes in WebXPRT 4, including new web technologies that may work well in a browser benchmark. Today, we’re going to focus on one of those technologies, WebAssembly, in more detail.

WebAssembly (WASM) is a binary instruction format that works across all modern browsers. WASM provides a sandboxed environment that operates at native speeds and takes advantage of common hardware specs across platforms. WASM’s capabilities offer web developers a great deal of flexibility for running complex client applications in the browser. That level of flexibility may enable workload scenario options for WebXPRT 4 such as gaming, video editing, VR, virtual machines, and image recognition. We’re excited about those possibilities, but it remains to be seen which WASM use cases will meet the criteria we look for when considering new WebXPRT workloads, such as relevancy to real life, consistency and replicability, and the broadest-possible level of cross-browser support.

One WASM workload that we’re investigating is a web-based machine learning workload with TensorFlow for JavaScript (TensorFlow.js). TensorFlow.js offers pre-trained models for a wide variety of tasks, including image classification, object detection, sentence encoding, and natural language processing. TensorFlow.js originally used WebGL technology on the back end, but now it’s possible to run the workload using WASM. We could also use this technology to enhance one of WebXPRT’s existing AI-themed workloads, such as Organize Album using AI or Encrypt Notes and OCR Scan.

We’re can’t yet say that a WASM workload will definitely appear in WebXPRT 4, but the technology is promising. Do you have any experience with WASM, or ideas for WASM workloads? There’s still time for you to help shape the future of WebXPRT 4, so let us know what you think!

Justin

Posted in benchmark, BenchmarkXPRT, Browser-based benchmarks, Collaborative benchmark development, Cross-platform benchmarks, image processing, Machine learning, Performance benchmarking, TensorFlow, Virtual reality, Web-based testing, WebAssembly, WebGL, WebXPRT, WebXPRT 3 | Tagged , , , , , , , , , |

Understanding the basics of AIXPRT precision settings

Justin Greene
By Justin Greene
on September 5, 2019

A few weeks ago, we discussed one of AIXPRT’s key configuration variables, batch size. Today, we’re discussing another key variable: the level of precision. In the context of machine learning (ML) inference, the level of precision refers to the computer number format (FP32, FP16, or INT8) representing the weights (parameters) a network model uses when performing the calculations necessary for inference tasks.

Higher levels of precision for inference tasks help decrease the number of false positives and false negatives, but they can increase the amount of time, memory bandwidth, and computational power necessary to achieve accurate results. Lower levels of precision typically (but not always) enable the model to process inputs more quickly while using less memory and processing power, but they can allow a degree of inaccuracy that is unacceptable for certain real-world applications.

For example, a high level of precision may be appropriate for computer vision applications in the medical field, where the benefits of hyper-accurate object detection and classification far outweigh the benefit of saving a few milliseconds. On the other hand, a low level of precision may work well for vision-based sensors in the security industry, where alert time is critical and monitors simply need to know if an animal or a human triggered a motion-activated camera.

FP32, FP16, and INT8

In AIXPRT, we can instruct the network models to use FP32, FP16, or INT8 levels of precision:

  • FP32 refers to single-precision (32-bit) floating point format, a number format that can represent an enormous range of values with a high degree of mathematical precision. Most CPUs and GPUs handle 32-bit floating point operations very efficiently, and many programs that use neural networks, including AIXPRT, use FP32 precision by default.
  • FP16 refers to half-precision (16-bit) floating point format, a number format that uses half the number of bits as FP32 to represent a model’s parameters. FP16 is a lower level of precision than FP32, but it still provides a great enough numerical range to successfully perform many inference tasks. FP16 often requires less time than FP32, and uses less memory.
  • INT8 refers to the 8-bit integer data type. INT8 data is better suited for certain types of calculations than floating point data, but it has a relatively small numeric range compared to FP16 or FP32. Depending on the model, INT8 precision can significantly improve latency and throughput, but there may be a loss of accuracy. INT8 precision does not always trade accuracy for speed, however. Researchers have shown that a process called quantization (i.e., approximating continuous values with discrete counterparts) can enable some networks, such as ResNet-50, to run INT8 precision without any significant loss of accuracy.

Configuring precision in AIXPRT

The screenshot below shows part of a sample config file, the same sample file we used for our batch size discussion. The value in the “precision” row indicates the precision setting. This test configuration would run tests using INT8. To change the precision, a tester simply replaces that value with “fp32” or “fp16” and saves the changes.

Config_snip

Note that while decreasing the precision from FP32 to FP16 or INT8 often results in larger throughput numbers and faster inference speeds overall, this is not always the case. Many other factors can affect ML performance, including (but not limited to) the complexity of the model, the presence of specific ML optimizations for the hardware under test, and any inherent limitations of the target CPU or GPU.

As with most AI-related topics, the details of model precision are extremely complex, and it’s a hot topic in cutting edge AI research. You don’t have to be an expert, however, to understand how changing the level of precision can affect AIXPRT test results. We hope that today’s discussion helped to make the basics of precision a little clearer. If you have any questions or comments, please feel free to contact us.

Justin

Posted in AI, AIXPRT, Benchmark metrics, Benchmarking, computer vision, image processing, Machine learning, object detection, ResNet-50 | Tagged , , , , , , , , , |

Check out the other XPRTs:

Forgot your password?
© Principled Technologies, Inc. 2025   |   Privacy policy

Cookie Notice: Our website uses cookies to deliver a smooth experience by storing logins and saving user information. By continuing to use our site, you agree with our usage of cookies as our privacy policy outlines.

x