Personal Computer Results

These are my personal computers that I’ve been benchmarking using Geekbench AI. This benchmarking tool provides detailed insights into how different systems perform AI-centric tasks, including image processing, object detection, and style transfer. It offers performance scores based on three types of precision: Single Precision, Half Precision, and Quantized. Each system’s GPU and CPU capabilities are tested, and the results help evaluate their strengths in handling various AI workloads.

The Benchmark Tests

Windows Tests

Windows CPU Benchmarks (ONNX Framework)

In these tests, I focused on the ONNX framework using the CPU backend on my Windows systems. The ONNX framework provides flexibility for deep learning models, and this test simulates the performance of CPU-bound AI tasks.
Windows CPU Benchmarks (OpenVINO Framework)

Another test using the OpenVINO framework and the CPU backend on Windows systems. OpenVINO is optimized for Intel hardware, making it a strong choice for performance in inference tasks. This test helps demonstrate the effectiveness of Intel CPUs in AI workloads.
Windows GPU Benchmarks (ONNX Framework + DirectML)

This test uses the ONNX framework with DirectML backend, leveraging the GPU for AI tasks. GPU tests are crucial for showing how well a system handles parallel processing in tasks like image recognition and machine learning model inference.

macOS Tests

macOS CPU Benchmarks (Core ML Framework)

Here, I tested the Apple M3 Max using Core ML with the CPU backend. Core ML is Apple’s machine learning framework optimized for macOS, and these tests show how well Apple silicon performs when executing AI workloads directly on the CPU.
macOS Neural Engine Benchmarks (Core ML Framework)

I also benchmarked Apple’s Neural Engine using Core ML. The Neural Engine is optimized for fast AI inference, and this test showcases its ability to perform AI tasks efficiently, particularly in tasks like object detection and style transfer.
macOS GPU Benchmarks (Core ML Framework)

Lastly, I tested the Apple M3 Max’s GPU using Core ML. This benchmark shows how Apple’s GPU handles AI workloads that benefit from parallel processing, particularly tasks like image super-resolution and pose estimation.

Understanding the Tests

The Geekbench AI benchmark tests systems on various real-world AI workloads. Here’s a breakdown of the types of tests and why they’re important:

Style Transfer: Transfers artistic styles between images, important for creative AI tasks like photo editing and art generation.
Object Detection: Detects and identifies objects in images and videos, used in applications like security, autonomous vehicles, and AR.
Pose Estimation: Estimates the positions of people or objects in images, crucial for motion tracking, fitness apps, and animation.
Image Super-Resolution: Enhances image resolution, used in fields like photography and medical imaging to improve visual clarity.
Image Classification: Categorizes images by identifying their contents, vital for tasks like facial recognition and photo organization.
Face Detection: Identifies faces within images or video, critical for security systems and user identification.
Depth Estimation: Measures the distance of objects in a scene, essential for 3D modeling and AR/VR.
Text Classification: Classifies blocks of text by type or topic, important in language processing tasks like spam detection.
Machine Translation: Translates text between languages, crucial for chatbots, document translation, and multilingual communication.
Image Segmentation: Splits an image into meaningful segments for analysis, widely used in medical imaging and autonomous driving.

Benchmark Results

This section provides detailed results from the Geekbench AI tests on my personal computers. The tables below show the Single Precision scores for each system. The systems are ranked based on their Single Precision Benchmark scores, providing an overview of their AI performance capabilities.

GPU Performance

Rank	System	RAM	GPU	VRAM	Single Precision Benchmark
1	MINISFORUM BD790i	96 GB DDR5	NVIDIA GeForce RTX 4070 Super	12 GB	27,830
2	ASUS Ryzen 7	32 GB DDR4	NVIDIA GeForce RTX 4070 Super	12 GB	25,495
3	ASUS ROG Zephyrus G16	32 GB DDR5	NVIDIA GeForce RTX 4090 Laptop	16 GB	22,486
4	ASUS ProArt P16	32 GB DDR5	NVIDIA GeForce RTX 4070 Laptop	8 GB	18,992
5	MacBook Pro M3 Max	36 GB Unified	Apple M3 Max 30-core GPU	Unified	13,411
6	Intel NUC9V7QNX	64 GB DDR4	NVIDIA GeForce RTX 4060	8 GB	13,112
7	ASUS ROG Zephyrus G15	16 GB DDR5	NVIDIA GeForce RTX 3060 Laptop	6 GB	12,858
8	MacBook Air M2	16 GB Unified	Apple M2 8-core GPU	Unified	5,419

ONNX CPU Performance

Rank	System	RAM	CPU	Single Precision Benchmark
1	MINISFORUM BD790i	96 GB DDR5	AMD Ryzen 9 7945HX	4,082
2	MacBook Pro M3 Max	36 GB Unified	Apple M3 Max	3,841
3	ASUS Ryzen 7	32 GB DDR4	AMD Ryzen 7 5800X	3,582
4	MacBook Air M2	16 GB Unified	Apple M2	3,214
5	ASUS ROG Zephyrus G16	32 GB DDR5	Intel Core Ultra 9 185H	3,076
6	ASUS ProArt P16	32 GB DDR5	AMD Ryzen AI 9 HX 370	2,983
7	ASUS ROG Zephyrus G15	16 GB DDR5	AMD Ryzen 9 6900HS	2,902
8	Intel NUC9V7QNX	64 GB DDR4	Intel Core i7-9850H	2,283

OpenVino CPU Performance

Rank	System	RAM	CPU	Single Precision Benchmark
1	MINISFORUM BD790i	96 GB DDR5	AMD Ryzen 9 7945HX	8,157
2	ASUS ProArt P16	32 GB DDR5	AMD Ryzen AI 9 HX 370	5,315
3	ASUS Ryzen 7	32 GB DDR4	AMD Ryzen 7 5800X	4,991
4	ASUS ROG Zephyrus G15	16 GB DDR5	AMD Ryzen 9 6900HS	4,218
5	MacBook Pro M3 Max	36 GB Unified	Apple M3 Max	3,841
6	Intel NUC9V7QNX	64 GB DDR4	Intel Core i7-9850H	3,434
7	MacBook Air M2	16 GB Unified	Apple M2	3,214
8	ASUS ROG Zephyrus G16	32 GB DDR5	Intel Core Ultra 9 185H	3,009

To submit your own benchmarks, clone the repository, create a new branch, and run your tests using Geekbench AI or another benchmarking tool. Add your system’s results using the template provided, including details like CPU, GPU, RAM, and your Single Precision, Half Precision, and Quantized scores. After committing your changes, push the branch to GitHub and open a pull request with a clear description of your submission.

[System Name] | [Framework] | [Backend] Benchmarks

[Provide a brief summary of your system’s performance. Describe the strengths and performance highlights from the tests.]

System Specs

CPU: [Name]
Motherboard: [Name]
RAM: [Size and Type]
GPU: [Name and VRAM]

Total Price: [Current Price]

Benchmark Scores

Single Precision: [Score]
Half Precision: [Score]
Quantized: [Score]

Top AI Tasks

AI Task	Performance (IPS)	AI Task	Performance (IPS)
1. [Task Name]	[Score]	4. [Task Name]	[Score]
2. [Task Name]	[Score]	5. [Task Name]	[Score]
3. [Task Name]	[Score]	6. [Task Name]	[Score]

[System Name] + [GPU Name] Performance Overview

[Provide a 3-4 sentence high-level summary of the system’s performance across both GPU and CPU tests. Highlight key strengths in AI tasks like model training, inference, or image processing. Mention any standout areas of performance, whether GPU-accelerated tasks or CPU efficiency. Keep it brief and general to reflect the overall benchmark results.]

Hardware Specs

CPU: [CPU Name]
Motherboard: [Motherboard Name] (Optional)
RAM: [RAM Size and Type]
GPU: [GPU Name and VRAM]

Estimated Total Price (2024): $[Current Price]

This reflects the current price for the system when purchased in 2024.

Price-to-Performance: [Calculated Ratio] IPS per $100

The [System Name] achieves a price-to-performance ratio of [Calculated Ratio] IPS per $100, based on its overall single precision GPU benchmark score of [Single Precision Score]. This summary should be 3-4 sentences, focusing on the system’s balance between cost and performance. Highlight how well it handles AI workloads and mention whether the performance is skewed towards GPU-heavy tasks or balanced across GPU and CPU tasks.

Price-to-Performance Calculation:
The price-to-performance ratio is calculated by dividing the single precision GPU benchmark score by the system price, then multiplying by 100.
Calculation: ([Single Precision Score] / [System Price]) × 100 ≈ [Calculated Ratio] IPS per $100