Synthetic benchmarks have been bleeding developer trust for years, often acting as mere marketing fodder for hardware giants to flaunt unrealistic performance metrics. Primate Labs is finally trying to stop the bleeding. On July 23, 2026, the company officially launched Geekbench 7, introducing a fundamental and desperately needed redesign to its widely used cross-platform testing software.
This major release directly addresses the industry's loudest complaints about synthetic testing. It deploys a smarter multi-core benchmark, expands machine learning workloads, and brings native support for NVIDIA's long-overdue CUDA API. Instead of blindly maxing out all available threads, Geekbench 7 completely rethinks processor evaluation by restricting multi-threaded testing strictly to workloads that actually utilize multiple cores in the real world.
A Reality Check for Multi-Core Scores
Geekbench 7’s biggest structural fix targets the multi-core illusion. Historically, benchmarking tools forced all available CPU cores to engage during a test. This artificially inflated scores for high-core-count processors, even if the simulated task normally relied on a single thread.
Primate Labs has finally eliminated this distortion. In Geekbench 7, a workload only runs in multi-threaded mode if the real-world application it models does the exact same thing.
For instance, the HTML5 Browser test is now excluded from the multi-threaded suite entirely, as modern web browsers operate on single or lightly threaded architectures. This architectural shift means that the resulting multi-core score provides a much more honest measure of how a device handles daily computing tasks.
Machine Learning and CUDA Integration
GPU performance testing receives a massive modernization in this release, acknowledging the dominant role of content creation and artificial intelligence in consumer software. After years of users begging for it, NVIDIA's CUDA finally joins OpenCL, Vulkan, and Metal as a supported API. This allows users to measure NVIDIA GPUs using the exact API that powers modern, high-demand professional applications, rather than relying on clunky workarounds.
The refreshed GPU benchmark explicitly targets machine learning capabilities by introducing distinct workloads that model everyday AI tasks. These tests include tracking faces and applying real-time filters for social media applications. It also tests upscaling images via machine learning to model super-resolution features, and blurring video backgrounds to simulate virtual conferencing tools.
Moving beyond the AI hype, the GPU suite dives deep into advanced image editing and synthesis workloads. Creators can benchmark hardware against RAW image processing, LUT-based video color grading, path tracing, and complex fluid simulations.
Heavier Media Workloads and Datasets
Modern hardware routinely chews through basic tasks without breaking a sweat, prompting Geekbench 7 to introduce heavier datasets and entirely new media workloads. The benchmark now actively measures how efficiently a CPU manages audio and video encoding, decoding, and live processing under genuine duress.
These new media workloads are highly specific to current computing trends. The software tests AV1 codec encoding to model modern screen-sharing applications, and uses the Opus codec to compress spoken-word audio simulating podcast tools. It also leverages the Whisper speech recognition model to decode video while generating live captions.
Pushing past standard media formatting, the developers integrated a new Game Physics workload built on the Jolt Physics engine. Photography benchmarks also see substantial upgrades, with the Photo Editor workload gaining a richer set of edits and the Photo Library workload adding support for importing modern formats like JPEG XL and DNG.
Standard productivity tests haven't been ignored, as they now force devices to process significantly larger and more complex files. The File Compression workload utilizes more varied archives containing source code and object code, while the PDF Viewer test forces systems to render complex park maps and technical academic papers.
Ultimately, Geekbench 7 feels like a necessary course correction for the entire benchmarking industry. By anchoring synthetic tests in reality and stripping away multi-core score inflation, Primate Labs has built a tool that actually reflects how we use our computers today.