It has been more than a year since the concept of AI PCs was hyped up, but it seems to have made a lot of noise with little impact. The market and consumers don’t seem to be buying into it either. Are AI PCs really “AI”? What truly defines a real AI PC? Let’s see what the real AI giants have to say.
1. The Rise of the AI PC Concept
AI PC, short for Artificial Intelligence Personal Computer, was first introduced by Intel in September 2023. In a short period, it gained widespread industry attention. Despite its relatively brief development history, the industry generally sees AI PCs as a turning point for the PC sector. Canalys defines AI PCs as desktops and laptops equipped with dedicated AI chipsets or modules (such as NPUs) to handle AI workloads.
The year 2024 is widely regarded as the inaugural year for AI PC applications, with major companies launching their own AI-powered computers.
In early March, Apple released the AI PC MacBook Air. On March 18, Honor introduced its first AI PC, the MagicBook Pro 16. Shortly after, AMD Chair and CEO Lisa Su announced that the AMD Ryzen 8040 series AI PC processors had started shipping. On March 22, Microsoft unveiled the Surface AI PC. On April 11, Huawei launched the new MateBook X Pro, integrating its Pangu AI model for the first time. To some extent, the PC industry’s strong push toward AI integration has indeed shown progress.
In Q4 2024, AI PC shipments reached 15.4 million units, accounting for 23% of total PC shipments for the quarter. For the full year, AI PCs made up 17% of total PC shipments. Apple led the market with a 54% share, followed by Lenovo and HP, each holding 12%. With the end of Windows 10 support driving a wave of device upgrades, AI PC market penetration is expected to continue rising in 2025.
But how much AI is really inside these so-called AI PCs?
2. AI PC: Big Hype, Little Impact
On February 23, 2024, Lenovo CEO Yang Yuanqing stated after the company’s latest financial report that global PC shipments were expected to grow by approximately 5% year-over-year in 2024. Despite some challenges, he firmly believes that AI will be a key driver in Lenovo’s business growth and transformation.
However, Yang also pointed out that the AI PC market is still in its early stages. While the hype is loud, actual sales and user adoption remain relatively low. He attributes this to factors such as technological maturity, user education, and market acceptance.
Many people are skeptical about the AI capabilities of current AI PC products. The main issue is that “AI” and “PC” (hardware) are essentially separate. Take Microsoft Copilot, the most significant AI application on PCs today. According to Intel and Microsoft’s joint definition of AI PCs, these devices must feature hybrid architecture chips, Copilot integration, and a dedicated physical Copilot button. However, in reality, any PC upgraded to the latest version of Windows 11 can use Copilot, as it relies entirely on Microsoft Azure’s cloud computing power rather than local PC hardware.
Meanwhile, NVIDIA, the undisputed leader in AI chip technology, has largely ignored Microsoft’s AI PC definition. And who has more authority in AI than NVIDIA?
NVIDIA has been building its AI ecosystem for years. Since its founding in 1993, it has been a pioneer in accelerated computing, boasting the most extensive CUDA ecosystem for AI productivity. High-performance PCs equipped with NVIDIA GPUs (discrete graphics cards) are far less dependent on OEM adaptations. They can run lightweight AI tools locally, such as large language models and basic Stable Diffusion image generation. Some can even handle mid-scale AI models, generating content much faster than integrated graphics.
So why is the AI PC market struggling to gain traction? Here are the key reasons:
✅ Insufficient NPU Computing Power in AI PCs
The AI performance of Intel’s NPU maxes out at 48 TOPS, while Intel Xe integrated graphics deliver around 28 TOPS. AI PCs with integrated graphics currently operate within the 10–45 TOPS range. In contrast, devices equipped with GeForce RTX 40-series GPUs—covering both laptops and desktops—offer options ranging from 200 to 1400 TOPS.
This year, NVIDIA’s newly released RTX 5090 GPU, built on the Blackwell architecture, marks a significant leap in performance. According to NVIDIA, the RTX 5090 delivers 4000 TOPS of AI computing power—three times that of its predecessor, the Ada Lovelace architecture. Compared to GPUs, the AI processing power of NPUs is negligible.
Even for mainstream AI applications, a single RTX 4080 or 4090 often struggles to provide abundant computing power locally. Given this, the limited performance of NPUs is even less useful.
✅ NPUs Lack DRAM and Cannot Independently Support Large Models
Today’s AI models heavily depend on DRAM. NPUs, by design, do not have dedicated DRAM and must rely on system RAM. Running large AI models with an NPU requires at least 64GB of DRAM. But if you’re already investing in more RAM, why not just use an APU or GPU instead? If you’re going to spend extra, it doesn’t matter which processor runs the AI workload.
Moreover, AI models are already optimized for APUs and GPUs, offering out-of-the-box compatibility with open-source software. In contrast, NPUs are not as well-supported.
✅ Limited Application Support for NPUs
In theory, NPUs can already handle LLMs (large language models), Stable Diffusion image generation, and common computer vision (CV) inference tasks like ResNet, YOLO, and Whisper speech-to-text. AI inference workloads are essentially matrix operations, which NPUs can execute with lower power consumption.
However, in practice, the AI functions available on Windows laptops today are minimal. The primary use cases leveraging NPUs are Windows Studio Effects’ background blur and CapCut’s background removal—far from groundbreaking applications.
As of now, there are very few local programs that support NPUs. Most of the actual NPU-powered features feel like gimmicks rather than game-changers. The real driver of AI adoption has been tools like ChatGPT, which demonstrate AI’s ability to solve practical problems. For NPUs to be truly useful, they need to run LLMs efficiently—but AI PCs’ current NPUs simply aren’t capable of this.
Ultimately, whether a PC is labeled as an “AI PC” is irrelevant. What truly matters is whether it has an NVIDIA GPU, as that determines its real AI capabilities.
3. The “Real” AI PCs from Three Major Players
While several manufacturers have promoted AI PC products, most have been more about marketing hype than actual AI capabilities. Many so-called AI PCs merely include an NPU chip but lack the ability to run large models locally—they can neither train nor infer AI models effectively.
AI PCs have been widely marketed in laptops. However, no ultrabook today can be considered a high-computing-power AI workstation. Instead, traditional high-performance gaming laptops and desktops with powerful GPUs are the real AI productivity machines.
For truly capable AI PCs, one must look at companies that develop high-performance GPUs—such as NVIDIA and AMD. At CES 2024, AMD introduced the AI Max 300 Strix Halo, while NVIDIA CEO Jensen Huang unveiled Project DIGITS. Apple, with its Mac Pro, has also entered this space. These three products are true “desktop AI supercomputers” for local AI model deployment.
✅ AMD’s Strix Halo: A High-Performance APU for AI
AMD’s Strix Halo lineup includes two versions:
- Consumer-grade Strix Halo – Designed for high-performance gaming laptops.
- Commercial-grade Strix Halo Pro – Targeted at mobile workstations.
Leaked 3DMark benchmark results show that the flagship Ryzen AI MAX+ 395 features:
- 16 Zen 5 CPU cores with 32 threads
- 40 RDNA 3.5 GPU cores (Radeon 8060S integrated graphics)
- Up to 120W TDP, three times that of a standard mobile APU
- Quad-channel LPDDR5X memory with a bandwidth of up to 256GB/s
Notably, the integrated Radeon 8060S GPU delivers over three times the performance of the previous Radeon 890M and is approaching the performance level of an RTX 4060 discrete GPU.
✅ NVIDIA’s Project DIGITS: The Smallest AI Supercomputer
NVIDIA calls Project DIGITS the “smallest AI supercomputer” to date. It features a custom GB10 superchip, which integrates:
- A Blackwell architecture-based GPU
- A Grace CPU, co-developed by NVIDIA, MediaTek, and ARM
The Blackwell GPU delivers 1 PFLOPS (1000 TFLOPS) of FP4 computing power, while the Grace CPU includes:
- 10 Cortex-X925 cores
- 10 Cortex-A725 cores
The GPU and CPU are connected via NVLINK-C2C, a chip-to-chip interconnect commonly used in supercomputers.
Additionally, Project DIGITS includes an NVIDIA ConnectX networking chip, enabling the GB10 superchip’s GPU to support multiple interconnect standards, such as:
- NCCL (NVIDIA Collective Communications Library)
- RDMA (Remote Direct Memory Access)
- GPUDirect, allowing direct access from AI applications
✅ Apple’s AI Advancements with M-Series Chips
In 2023, Apple introduced the M3 series chips, featuring next-gen GPUs with the biggest leap in Apple GPU architecture history:
- 2.5x faster rendering than M1 chips
- New Dynamic Caching technology
- Hardware-accelerated ray tracing and mesh shading
Apple also introduced unified memory architecture (UMA), allowing up to 128GB of unified memory. This enables AI developers to work with large Transformer models containing billions of parameters, something previously impossible on laptops.
Apple later released the M4 Pro, claiming it outperforms traditional AI PC chips.
✅ The Power of Unified Memory Architecture
All three of these AI computing solutions—AMD Strix Halo, NVIDIA Project DIGITS, and Apple’s M-series—adopt Unified Memory Architecture (UMA). The benefits of UMA include:
- Combining system RAM and GPU VRAM into a single memory pool
- Reducing data transfer bottlenecks between the CPU and GPU
- Eliminating the need for memory copying, speeding up AI workloads
- Enabling larger memory capacities, addressing GPU VRAM limitations in consumer-grade hardware
Interestingly, Apple’s M1 chip was the first to introduce a unified memory design, not NVIDIA.
4. DeepSeek Ignites the Desktop AI Supercomputer Battle
Recently, the severe shortage of DeepSeek’s online computing resources has driven a surge in demand for local deployment of large AI models. In response, the three major “real AI PC” manufacturers have started integrating DeepSeek into their systems.
DeepSeek, as an MoE (Mixture of Experts) model, has high VRAM requirements but relatively lower demands on raw computing power and memory bandwidth. This plays directly into the strengths of desktop AI supercomputers equipped with unified memory architecture, which provides significantly larger memory pools.
Earlier, an AI enthusiast successfully deployed DeepSeek V3 on eight M4 Pro Mac minis. Similarly, it is expected that four Project DIGITS units could also be used to run DeepSeek V3—with significantly faster generation speeds.
According to AMD, Strix Halo APUs can deploy a 70B model 2.2 times faster than an RTX 4090 while consuming 87% less power. Some users are already planning to switch to Strix Halo-powered laptops once they hit the market, seeing local AI model deployment as an exciting new frontier. There is even speculation that within a few years, it may be possible to locally run 671B INT8 or FP8 models.
Beyond AI model deployment, increased RAM and CPU performance also translate to improved general computing performance across other workloads.
5. AI PC: A Story with Many Narratives
The concept of AI PCs has been molded differently by different manufacturers—each telling their own version of the story. OEMs are investing heavily in local AI applications, with some software running both locally and in the cloud, while cloud services integrate domestic AI models for commercialization. This hybrid model of local AI and cloud AI could prove to be a lucrative opportunity.
Key factors that could drive the mass adoption of AI-powered PCs include:
- Low latency + enhanced privacy protection
- AI applications such as GPT-like chat models, Stable Diffusion image generation, voice cloning, AI frame interpolation, background removal, and image inpainting
- AI PCs with powerful edge computing capabilities, large memory (VRAM), and highly optimized AI software
6. Conclusion: AI PCs Are More Than Just Hype
AI PCs are not just a marketing gimmick. Whether it’s more accessible AI, more energy-efficient AI, or higher-performance AI, the industry is actively advancing technologies and exploring new markets. The convergence of cloud and local AI computing is making AI more seamless and pushing AI PCs closer to real-world applications.
Source: Internet
Related:
- AI PCs Unveiled: Do Everyday Users Need to Upgrade?
- What is Windows AI PC? Hardware & Software Needs
Disclaimer:
- This channel does not make any representations or warranties regarding the availability, accuracy, timeliness, effectiveness, or completeness of any information posted. It hereby disclaims any liability or consequences arising from the use of the information.
- This channel is non-commercial and non-profit. The re-posted content does not signify endorsement of its views or responsibility for its authenticity. It does not intend to constitute any other guidance. This channel is not liable for any inaccuracies or errors in the re-posted or published information, directly or indirectly.
- Some data, materials, text, images, etc., used in this channel are sourced from the internet, and all reposts are duly credited to their sources. If you discover any work that infringes on your intellectual property rights or personal legal interests, please contact us, and we will promptly modify or remove it.