Introducing Computer Vision 2.0, Our Next‑Generation AI Hardware Benchmark

Today marks the launch of Computer Vision 2.0, our next‑generation computer‑vision benchmark built to evaluate modern artificial intelligence (AI)‑capable hardware with accuracy, fairness and real‑world relevance. Hardware has advanced at an unprecedented pace, with new architectures and NPUs emerging every year.Gartner expects AI PCs to become standard by 2029. 

Computer Vision 2.0 introduces new models and standardized execution paths. Our realistic computer vision workload designs provide clarity and make results easier to understand. We focus on everyday workflows, from background blurs in video calls to image grouping, object detection, screenshot descriptions and accessibility enhancements.

Key features of this release include:

Modern transformer‑based model suite – we’ve replaced older CNN‑only architectures with cutting‑edge transformer models. DETR for detection, SAM2 for segmentation, BLIP for captioning, ESRGAN for image super-resolution, and ConvNeXt for image classification. These models better represent the workloads used in modern applications.

A standardized Windows ML execution for fair comparisons – take the fatigue out of understanding which tools you need to benchmark each AI device. We provide a single, standardized score for evaluating all AI‑capable Windows hardware, including NPUs. Selecting the correct precision (FP32/FP16/INT8) traditionally requires deep AI expertise. Computer Vision 2.0 removes that barrier with auto‑precision selection, choosing the optimal setting per device to support consistency, fairness and ease of use.

Cross‑platform support from day one – Computer Vision 2.0 launches with macOS support on day one, enabling cross‑platform comparisons for organizations with mixed fleets.

Realistic workload designs and accurate thresholds for integrity – our workload performance measurements reflect real employee experiences, and this latest release now includes scenarios that mirror true multimodal use. Examples of this include the ability to classify and group images locally on photo applications, and camera tools to detect objects, annotate, and apply background effects.

The work we do on our Procyon AI benchmarks enables our customers to make more confident, data‑driven procurement decisions aligned with the future of AI PCs. For more information about Computer Vision 2.0 or to discuss how it can support your hardware evaluation strategy, contact us today.

Read more on the UL Benchmarks website