OH-Tech

Why is the MacBook getting more expensive? An explanation of the costs of Unified Memory and on-device AI

OH-TECH7 min read
Why is the MacBook getting more expensive? An explanation of the costs of Unified Memory and on-device AI
MacBook prices have risen sharply since 2023, especially after the M3 generation, not only because of more powerful chips but also because of a core reason: Apple Intelligence and Local AI require enormous amounts of RAM. Unified Memory on Apple Silicon is the most expensive component in every new MacBook, and the currently available 8GB upgrade is priced disproportionately to the actual cost of the components.

Why have MacBook prices increased from 2023 to now?

MacBook prices have steadily increased across the M1, M2, M3, and M4 generations. The main reason lies not in the processor but in unified memory (Unified Memory). Apple charges around 4,5 million VND for the 8GB RAM upgrade in Vietnam — equivalent to the price of a mid-range smartphone — while the actual retail market cost of an LPDDR5 DRAM component is only around 1-1,5 million VND for 8GB. The reason lies in the distinctive memory architecture of Apple Silicon.

What is Unified Memory, and why is it more expensive than regular RAM?

Unified Memory is an architecture in which the CPU, GPU, Neural Engine, and ISP share a single physical memory pool, instead of having separate memory for each component as in traditional computers. This design provides extremely high bandwidth and significantly reduces latency, while allowing AI processing cores (Neural Engine) to access data directly without copying it back and forth. However, this architecture requires an LPDDR5 DRAM chip to be placed directly on the SoC package via chip-on-substrate technology or PoP (Package on Package). The manufacturing process is more complex, the yield rate is lower, and Apple is the largest-volume customer of Samsung, SK Hynix, and Micron for the mobile LPDDR5X lineup — this gives Apple reverse bargaining leverage: they can push retail prices higher without necessarily having to buy more cheaply.

Why does local AI drive a surge in demand for MacBook RAM?

Apple Intelligence requires a large amount of memory because large language models (LLMs) and generative AI models run entirely on-device, not in the cloud. Below are the actual memory requirements for Apple Intelligence on Apple Silicon:
  • 8GB RAM: Can run basic Apple Intelligence, but memory is nearly full when many apps are open at the same time. Swapping to the SSD occurs frequently, causing slowdowns.
  • 16GB RAM: The minimum recommended amount for running Apple Intelligence comfortably, sufficient for most local AI tasks without lag.
  • 24GB RAM: Suited to professional users, allowing local AI to run alongside demanding applications such as Final Cut Pro, Logic Pro, or many browser tabs.
  • 36-128GB RAM (M3 Max, M4 Max): Designed for developers running local AI models, fine-tuning models, or on-site machine learning tasks.
Every billion parameters (parameter count) of an AI model requires a proportional amount of RAM. A model with approximately 7 billion parameters (such as those commonly used in Apple Intelligence) requires a minimum of 14-16GB RAM just to load the model, not to mention the context buffer and data from running applications.

How does memory bandwidth affect AI performance?

Memory bandwidth is the data transfer speed between the SoC and DRAM, measured in GB/s. For on-device AI, the higher the bandwidth, the faster the AI model’s weights can be transferred between memory and the Neural Engine, reducing inference latency. The actual bandwidth specifications of the Apple Silicon lineup:
  • M4 (10-core CPU): 120 GB/s memory bandwidth
  • M4 Pro: 273 GB/s
  • M4 Max: 546 GB/s (previous-generation memory bandwidth); the new-generation M4 Max models achieve higher bandwidth
  • M3 Max: Up to 800 GB/s
The 800 GB/s figure of the M3 Max far exceeds that of a mid-range discrete GPU (typically around 400-600 GB/s over GDDR6). This is why a MacBook Pro M3 Max with 128GB RAM can run a 70-billion-parameter AI model locally without a discrete GPU. This speed comes from integrating DRAM directly into the SoC package, with a corresponding increase in manufacturing costs.

How much is a reasonable upgrade to 8GB RAM on a MacBook?

Currently, Apple charges for upgrading 8GB Unified Memory as follows in the Vietnamese market:
  • From 8GB to 16GB: approximately 4,5 million VND
  • From 16GB to 24GB: approximately 4,5 million VND
  • From 24GB to 32GB: approximately 4,5 million VND
The retail price of a 16GB LPDDR5X DRAM module is currently around 1,8-2,2 million VND. The price difference reflects R&D costs, specialized packaging, Apple's profit margin, and especially the cost of optimizing software for each specific memory configuration on macOS.

Conclusion: Should You Buy a MacBook with More RAM?

If you plan to use Apple Intelligence, run a local AI model (via Ollama, LM Studio, or other generative AI apps), or do professional content creation, investing in more RAM is sensible and will extend the device's usable lifespan. If you only use it for office tasks, browsing the web, and watching movies, 16GB is enough, but think carefully before choosing an 8GB configuration because Apple Intelligence and future AI apps will require increasingly more memory. The increase in MacBook prices is not only Apple’s pricing strategy but also reflects the real demand for memory from local AI. As AI chips (NPUs) in Windows PCs are also pushing RAM requirements to 16GB and 32GB as the new standard, this trend will continue for years to come.

Frequently asked questions

Can Apple Intelligence run on a MacBook with 8GB RAM?

8GB RAM can run basic Apple Intelligence, but memory will be nearly full when additional apps are opened. Apple recommends 16GB RAM for a smooth Apple Intelligence experience without slowdowns due to swapping.

Why does upgrading RAM on a MacBook cost much more than RAM on a PC?

Unified Memory on Apple Silicon is placed directly on the SoC package using a proprietary packaging technology (PoP/chip-on-substrate), rather than a conventional standalone DDR5 RAM module. The manufacturing process is more complex, the number of suppliers is limited (only Samsung, SK Hynix, Micron), and the cost of optimizing the software for each configuration also drives up the price.

Which MacBook is suitable for running local AI (Local AI)?

A MacBook with an M3 Pro/Max or M4 Pro/Max chip and at least 24GB RAM is a suitable choice for running local AI models such as Llama 3 and Mistral. An M4 Max with 128GB RAM can run models of up to 70 billion parameters directly on the device.

How does memory bandwidth (Memory Bandwidth) affect AI performance?

The higher the bandwidth, the faster the Neural Engine can access AI model weights, reducing latency during inference. The M3 Max achieves 800 GB/s — much higher than many mid-range GPUs using GDDR6 — thanks to its directly integrated Unified Memory architecture.

How is Unified Memory different from regular RAM?

Regular RAM (DDR5) is dedicated memory for the CPU, while the GPU has its own dedicated memory (VRAM). Unified Memory uses a shared memory pool for all processing units (CPU, GPU, Neural Engine, ISP), eliminating the need to copy data back and forth, significantly increasing bandwidth and reducing latency.