This is a big year. Microsoft discontinued technical support for any computers running Windows 10, and AI adoption is a continuing trend across businesses. By investing in Windows 11 Pro AI PCs with processors that contain integrated neural processing unit (NPU) architecture as well as the usual CPU and GPU architecture, your company opens the door to a world of possibilities. NPUs are designed to accelerate AI inference, computer vision, large language model (LLM), and machine learning (ML) workloads.

To help you decide which AI PC is right for you, we compared productivity and on-device AI performance metrics on these desktops:

• HP EliteDesk 8 Mini G1a Desktop Next Gen AI PC powered by an AMD Ryzen^™ AI 7 PRO 350 processor
• Dell Pro Micro Plus Desktop powered by an Intel^® Core^™ Ultra 7 265 vPro^® processor
• Lenovo ThinkCentre M90q Gen 6 powered by an Intel Core Ultra 7 265T vPro processor

As AI use cases continue to unfold, system performance, especially for on-device AI workloads, is becoming more and more important. We found the AMD Ryzen^™ AI 7 Pro 350 processor-powered HP EliteDesk 8 Mini G1a Desktop Next Gen AI PC is equipped to help your organization keep up in a rapidly changing business environment.

Our testing

All three Windows 11 Pro AI PCs we evaluated provided built-in AI capabilities, enhanced built-in Windows securities to plug potential pre-Windows 11 vulnerabilities¹, and NPU technology:

*The results we report reflect the specific configurations we tested. Any difference in the configurations—as well as screen brightness, network traffic, and software additions—can affect these results. For a deeper dive into our testing parameters and procedures, see the science behind the report.

To measure productivity and on-device AI performance, we ran these benchmark tests:

• Geekbench AI
• LM Studio
• MLPerf Client Benchmark
• PassMark PerformanceTest 11
• Procyon^® AI Computer Vision Benchmark

We also measured noise output while the desktops ran a sustained Cinebench 2024 workload.

Note: The graphs in this report use different scales to keep a consistent size. Please be mindful of each graph’s data range as you compare.

Better cutting-edge and everyday performance

AI insights are transforming industries. Don’t be left behind. By prioritizing on-device AI performance suited to your specific needs, you’re investing in your company’s potential. Getting answers in less time, while keeping control of your sensitive data on-device, is the first step on your road to success. But there’s no one test that paints a broad picture of on-device AI and general productivity performance. That’s why we examined CPU, GPU, and NPU performance from multiple angles.

Cutting-edge capabilities

To measure GPU performance for on-device ML workloads, we ran the Geekbench AI benchmark. This benchmark used real-world ML apps to provide a multidimensional picture of on-device AI performance.⁵ The precision level scores below reflect different AI model requirements: Full Precision (FP32) is the most accurate and the most resource-intensive, Half Precision (FP16) is less accurate but more efficient, and Quantized (INT8) is the most resource-efficient and least accurate of all.⁶ For this evaluation, we used the Open Neural Network Exchange (ONNX) open-source AI framework and DirectML AI backend for ML on Windows.

We found the AMD Ryzen^™ AI 7 PRO 350 processor-powered HP EliteDesk 8 Mini G1a Desktop Next Gen AI PC outperformed both Intel Core Ultra 7 processor-powered AI PCs.

To measure CPU performance for on-device AI Chat models, we ran LM Studio, which uses local LLMs to capture token metrics.⁸ According to Microsoft, LLM tokens are “words, character sets, or combinations of words and punctuation…”⁹ In these tests, we ran the Meta-Llama-3.1-8B-Instruct-Q4_K_M model. This LLM predicts what the person typing the query is going to type next based on what they’ve already input.

When processing the local LLM, the HP EliteDesk 8 Mini G1a desktop produced the first token and processed the data in significantly less time than the competitors. The less time it takes an on-device AI chatbot to accurately figure out what you’re looking for and produce a result, the less frustrating the user experience.

To measure NPU performance for AI inference engine models, we used the UL Procyon AI Computer Vision Benchmark.¹⁰ In our integer-optimized (INT8) testing, we used the API optimized for each system’s NPU: the AMD Ryzen^™ AI API on the AMD processor-based system and the Intel OpenVINO^™ inference API on the Intel processor-based systems. The individual AI inference tasks and their use cases were:

• MobileNetV3, ResNet-50, and Inception-v4: Convolutional neural networks (CNNs) widely used for image recognition, object detection, and image classification tasks. Essential for research institutions, tech companies, and individuals.^11,12,13
• DeepLabv3 and YOLOv3: Deep neural networks (DNN) that distinguish between different objects and features within images and videos. Used by healthcare providers, manufacturers, and video surveillance companies.^14,15
• Real-ESRGAN: A generator and discriminator network (GAN) that enhances image quality and resolution. Used by digital artists, medical professionals, and real estate firms.^16,17

These results show the large gap between the overall scores for each AI PC. To see the average wait times for each inference engine on each AI PC, see the science behind the report.

To measure NPU performance for on-device GenAI tasks, we used the MLPerf Client Benchmark to perform text and speech generation tasks. MLCommons says, this benchmark “provides clear metrics for understanding how well systems handle generative AI workloads.”²⁰ Because each language-based workload has its own ideal use case, we tested performance with three LLMs:

• Llama 2 7B Chat: This chatbot-based model is fine-tuned for conversational dialog and instruction use cases²¹ and used in AI-based document summarization and question answering scenarios.
• Llama 3.1 8B Instruct: This natural language processing (NLP) model is optimized for multilingual dialog and conversation use cases.²²
• Phi 3.5 Mini Instruct: This NLP model is designed to accelerate long-context tasks, including meeting summarization, long document summarization and QA, and other document-based retrieval tasks.²³

For these tests, we used the vendor-optimized execution path for each platform: ONNX Runtime GenAI with the AMD Ryzen^™ AI SDK in hybrid mode (NPU + iGPU) on the AMD processor-based system and the Intel OpenVINO inference API targeting the Intel NPU on the Intel processor-based systems.

Once again, we found the AMD Ryzen^™ AI 7 PRO 350 processor-powered HP EliteDesk 8 Mini G1a desktop produced the first token and processed the data in significantly less time than the Intel Core Ultra 7 processor-powered AI PCs. The less time it takes GenAI to summarize a doc or translate dialog, the quicker the user can get up to speed or understand the assignment.

Day-to-day capabilities

We also wanted to see how the AI PCs performed average everyday tasks. To do that, we ran the PassMark PerformanceTest 11 benchmark. PassMark PerformanceTest 11.0 runs CPU, video card, graphics hardware, and SSD tests and combines the results into an overall rating. This overall rating shows how the PCs handle complex mathematical calculations, 2D and 3D graphics, IOPS, and database operations.²⁴

Conclusion

In our hands-on tests, a HP EliteDesk 8 Mini G1a Desktop Next Gen AI PC with an AMD Ryzen^™ AI 7 PRO 350 processor outperformed a Dell Pro Micro Plus Desktop with an Intel Core Ultra 7 265 vPro processor and a Lenovo ThinkCentre M90q Gen 6 with an Intel Core Ultra 7 265T vPro processor. This compact desktop delivered cutting-edge AI capabilities alongside strong general computing power, making it a solid AI PC choice for businesses seeking to harness on-device AI for enhanced productivity and data privacy.