What Is Unified Memory? How Does Unified Memory Work? [MiniTool Wiki]

What Is Unified Memory?

Unified memory is about minimizing data redundancy that is copied between the different parts of memory used by the CPU, GPU, etc. Copying is slow and wastes memory capacity. In traditional memory implementations, part of the RAM is reserved for the GPU. If your laptop advertises 16GB of RAM and 2GB is allocated to the GPU, you only have 14GB available for system tasks. Apple solved this problem with UMA, making memory allocation smoother and improving performance.

Why Does Your Computer Need Unified Memory?

The Unified Memory Architecture on Apple's silicon introduces several changes to the memory system on the computer. From changing how RAM connects to computing units to redefining memory architecture, Apple is changing the way memory systems are designed to improve system efficiency.

That said, the new architecture creates a race condition between the CPU, GPU, and Neural Engine, increasing the amount of RAM required by the system.

How Does Unified Memory Work?

In the case of general-purpose systems, the RAM connects to the CPU using sockets on the motherboard. This connection limits the amount of data sent to the CPU.

Apple Silicon, on the other hand, uses the same substrate to house RAM and SoC. While RAM is not part of the SoC in this architecture, Apple uses an interposer substrate (Fabric) to connect the RAM to the SoC. An interposer is nothing more than a layer of silicon between the SoC and the RAM.

Compared to traditional sockets, which rely on wires to transmit data, interposers allow RAM to connect to chipsets using through-silicon vias. That means MacBooks with Apple's silicon technology embed RAM directly into the package, allowing for faster transfers of data between memory and the processor. RAM is also physically closer to where the data is needed (the processor), allowing data to get to where it needs to be faster.

Because of this difference in connecting RAM to the chipset, it can access high data bandwidth.

In addition to the differences mentioned above, Apple has also changed the way the CPU and GPU access the memory system.

As mentioned earlier, GPUs and CPUs have different memory pools in traditional setups. Apple does the opposite, allowing the GPU, CPU, and Neural Engine to access the same memory pool. Therefore, data does not need to be transferred from one memory system to another, further improving the efficiency of the system.

Due to all these differences in memory architecture, unified memory systems provide SoCs with high data bandwidth.

How Much Unified Memory Do You Need?

Now that we have a basic understanding of the unified memory architecture, we can see how much you need.

Although Unified Memory Architecture has several advantages, it still has some drawbacks. First, the RAM is connected to the SoC, so the user cannot upgrade the RAM on the system. Also, the CPU, GPU, and Neural Engine access the same memory pool. As a result, the amount of memory required by the system increases dramatically.

So, if you surf the web a lot and use a lot of word processors, 8 GB of RAM should be more than enough. But if you use Adobe Creative Cloud programs a lot, getting the 16 GB variant is a better choice as you'll have a smoother experience editing photos, videos, and graphics on your machine.

If you're training many deep learning models or working with video timelines with lots of layers and 4K footage, you should also consider the M1 Ultra with 128 GB of RAM.

  • Linkedin
  • Reddit