Primate Labs released Geekbench 7 on July 25, 2026, overhauling its cross-platform benchmark with a new score scale, expanded workloads, and—for the first time—native NVIDIA CUDA GPU testing. An early result from an Apple M5 Max recorded 3,730 single-core and 35,139 multi-core points, but the bigger story for Windows users is a benchmark that now mirrors how modern PCs actually work.

A New Scoring Baseline

The most immediate change: Geekbench 7 scores are not comparable to Geekbench 6. The benchmark introduces a fresh scale anchored to real hardware—a Lenovo Legion with an AMD Ryzen 7 7700 serves as the CPU baseline at 2,500 points, while the same system’s NVIDIA GeForce RTX 4060 anchors the GPU baseline at 100,000 points. A number that looks familiar no longer carries the same performance meaning.

What’s Inside the CPU Test Suite

Geekbench 7 replaces its predecessor’s workload mix with tasks picked for today’s interactive, media-heavy computing. Instead of narrow synthetic routines, the benchmark now includes:

  • File Compression: LZ4, zlib, and Zstandard across multiple archive types, with SHA-1 verification. This mirrors compression in developer tools, installers, and backups better than a single library.
  • PDF Viewer: Uses PDFium, the rendering engine from Chromium, making the result directly relevant to Windows users who spend hours in browsers and document viewers.
  • Photo Library: Adds JPEG XL and DNG processing with image tagging and database operations—more like a real photo app than a simple decode.
  • Media Workloads: A new dedicated media group brings video encoding (AOM AV1 on simulated screen sharing), audio encoding (Opus codec on music and speech), and a video decoder that chains AV1 unpacking, Opus decoding, resampling, and Whisper-powered caption generation. This reflects the multi-stage pipelines common in conferencing and streaming.
  • Headless browser rendering, Clang compilation, Jolt physics, HDR image operations, and more.

The suite intentionally avoids saturating all cores for long periods. It measures bursty, interactive performance—how quickly a system completes a sequence of common user tasks rather than sustained throughput.

Native CUDA Arrives for Windows Users

For the first time, Geekbench’s GPU test can talk directly to NVIDIA hardware via CUDA. Previous versions relied on OpenCL or Vulkan abstractions, which introduced overhead and didn’t always represent how creators and engineers use their GPUs. The new GPU workloads are chosen for modern visual and AI-accelerated tasks:

  • DeepLabV3+ for background segmentation and blur (think Teams or Zoom effects)
  • RetinaFace for face detection and filters
  • RFDN image upscaling from 256×256 to 1024×1024
  • Horizon correction, photo filters, LUT color grading, RAW processing with denoising and demosaicing
  • Feature matching and path tracing using the Blender BMW scene
  • Simulation workloads

Because CUDA is NVIDIA’s native compute API, Geekbench 7 can now produce scores that more accurately reflect how GeForce and RTX cards handle real GPU-accelerated applications. Cross-API comparisons—CUDA on NVIDIA, Metal on Apple silicon, OpenCL on AMD, Vulkan on Intel—remain tricky, but within the NVIDIA ecosystem, the benchmark has become far more relevant.

Interpreting the Early M5 Max Numbers

The first notable Geekbench 7 result comes from an 18-core Apple M5 Max with 48GB of unified memory running macOS 26.5.2. Its single-core score of 3,730 and multi-core of 35,139 are impressive, particularly in sub-tests like Video Player (5,357) and Structure from Motion (4,439). These numbers underscore Apple’s strength in latency-sensitive and media-focused tasks, but they are a single data point—not a product review. Cooling, power mode, and firmware differences will swing scores on both Macs and Windows machines.

For Windows users tracking the ARM PC transition, the M5 Max result shows the competitive pressure on Qualcomm’s Snapdragon X, Intel’s Core Ultra, and AMD’s Ryzen AI chips. However, a high single-core Geekbench score does not automatically translate to superior Windows productivity. Native vs. emulated software, driver maturity, and storage speed often matter more in daily use.

What This Means for Your Windows PC

Geekbench 7 resets the comparison clock. Any inventory of Geekbench 6 results becomes historical reference. If you rely on the benchmark to gauge system health, overclocking gains, or the impact of driver updates, you must run the new version.

The update also gives Windows users a more meaningful GPU yardstick. A CUDA-based score on an RTX 4090 laptop now reflects the same compute path used by Adobe Premiere Pro, Blender, and local machine learning tools. That’s a significant upgrade from generic compute tests.

But remember: Geekbench 7 is a short-burst, interactive benchmark. It won’t tell you how a workstation handles hour-long renders, large codebase compilations, or database transactions. For those, you still need Cinebench, Blender rendering tests, SPEC, or real application benchmarks.

How We Arrived at a Benchmark Overhaul

Geekbench has always aimed for cross-platform simplicity. When it launched, most PCs used homogenous multi-core CPUs and discrete GPUs primarily for gaming. The computing landscape now includes hybrid architectures (performance and efficiency cores), integrated NPUs, and GPUs that accelerate everything from video calls to AI inference. Geekbench 6’s workloads couldn’t capture these shifts. The new suite finally acknowledges that a modern PC’s value isn’t just in heavy multi-threaded number crushing but in fluid video playback, quick photo edits, and smart background effects.

Your Geekbench 7 Action Plan

To get the most out of Geekbench 7 on Windows, follow a disciplined approach:

  1. Download and run the benchmark at least three times. Consistency matters more than a single high score. Use Geekbench 7.0.0 for Windows.
  2. Document your exact configuration: Windows version, GPU driver version (e.g., Game Ready 560.xx), BIOS revision, memory speed, and power plan. This helps when comparing before/after changes.
  3. Test in both balanced and performance power modes on laptops, with AC power connected. Note if OEM control software alters fan curves or power limits.
  4. Separate CPU and GPU conclusions. A fast CPU score doesn’t predict GPU performance, and a high GPU score doesn’t guarantee a responsive desktop.
  5. Only compare within Geekbench 7. Never use version 6 scores to calculate performance gains or losses.
  6. Use real applications for critical decisions. Test games in actual games, video exports in your editor, and AI model inference in your framework. Geekbench 7 provides a useful health check, not a final verdict.
  7. Wait for the database to populate. A single M5 Max result is interesting, but patterns across many submissions reveal realistic performance bands. Check back after the Windows community uploads a range of systems.

Outlook

Geekbench 7 will improve as its result database grows and Primate Labs refines the early version. Account management and result browsing need polish, and the first submissions will inevitably invite targeted optimizations from hardware vendors. Over the coming months, expect to see scores from Ryzen 9000, Core Ultra 200, Snapdragon X Elite, and Arc Battlemage GPUs. The benchmark’s lasting value will depend on how well it predicts real-world performance on the tasks users actually care about—and the addition of native CUDA is a strong step in that direction for Windows.