Geekbench AI: Benchmarking NPU, GPU, and CPU Tensor Tasks

This article explores the critical role of Geekbench AI, formerly known as Geekbench ML, in standardizing the evaluation of artificial intelligence workloads across modern computer hardware. It breaks down how the benchmarking tool measures tensor task performance on Central Processing Units (CPUs), Graphics Processing Units (GPUs), and Neural Processing Units (NPUs). Furthermore, it explains how the software evaluates real-world machine learning operations through various data precision levels and software frameworks to give users an accurate picture of hardware capabilities.

The Evolution from Geekbench ML to Geekbench AI

As artificial intelligence moved from cloud servers to local consumer devices, the hardware industry needed a standardized way to measure on-device AI performance. Primate Labs rebranded Geekbench ML to Geekbench AI to reflect the broader scope of modern artificial intelligence workloads. The benchmark's primary role is to provide a cross-platform tool that evaluates how effectively different system architectures execute tensor mathematics—the foundational calculations behind deep learning and neural networks.

Evaluating the Three Hardware Pillars: CPU, GPU, and NPU

Modern system-on-chips (SoCs) and desktop computers utilize three distinct processors capable of running AI workloads. Geekbench AI isolates and tests each component to show where a system excels or bottlenecks.

  • CPUs (Central Processing Units): The benchmark tests the CPU to establish a baseline. While CPUs are not specialized for massive parallel tensor tasks, they handle lightweight, bursty AI workloads well due to their high clock speeds and immediate availability without waking up other system components.
  • GPUs (Graphics Processing Units): GPUs are built for parallel processing, making them naturally suited for the heavy matrix multiplication required in tensor tasks. Geekbench AI evaluates the GPU to measure peak throughput for demanding AI workloads, such as image generation and complex computer vision.
  • NPUs (Neural Processing Units): NPUs are dedicated AI accelerators designed specifically for sustained tensor math at a fraction of the power consumption of a GPU. Geekbench AI plays a vital role here, as NPUs are notoriously difficult to compare across brands like Apple, Intel, AMD, and Qualcomm. The benchmark runs identical workloads on these disparate architectures to provide a unified performance score.

Measuring Tensor Precision: FP32, FP16, and INT8

A processor's raw speed is only part of the AI performance equation; how it handles different data types is equally important. Tensor tasks rely on processing massive arrays of numbers, and developers often lower the precision of these numbers (a process called quantization) to increase speed and reduce memory usage. Geekbench AI evaluates hardware across three specific precision metrics:

  1. Single Precision (FP32): Uses 32-bit floating-point numbers. This tests the hardware's ability to handle maximum accuracy, which is often handled best by CPUs and high-end GPUs.
  2. Half Precision (FP16): Uses 16-bit floating-point numbers. This is the standard for many modern GPU and NPU workloads, balancing speed and accuracy.
  3. Quantized (INT8): Uses 8-bit integers. NPUs are heavily optimized for INT8 math. By testing INT8, Geekbench AI reveals how fast hardware can run compressed, highly efficient AI models designed for mobile devices and laptops.

Instead of combining these into a single, potentially misleading number, Geekbench AI outputs three separate scores for FP32, FP16, and INT8. This clearly demonstrates whether a chip relies on raw power or quantization shortcuts to achieve its speed.

Real-World Workloads and Software Frameworks

Hardware cannot execute tensor tasks without a software layer to translate the code. A major role of Geekbench AI is testing the synergy between the physical processor and its corresponding AI framework.

When benchmarking a device, Geekbench AI runs real-world computer vision tasks (like image segmentation and object detection) and natural language processing tasks (like text translation and classification). It executes these tasks using the native frameworks of the operating system, such as Apple’s Core ML, Google’s TensorFlow Lite, Intel's OpenVINO, or Microsoft’s ONNX Runtime.

By testing the hardware through these frameworks, Geekbench AI does not just measure theoretical silicon performance. Instead, it measures the actual performance a user or developer will experience when running AI-enhanced applications, exposing whether a manufacturer's software drivers are truly utilizing the underlying CPU, GPU, or NPU hardware to its fullest potential.