MaixPy MaixCAM Memory Usage Guide

Introduction to MaixPy MaixCAM Memory

MaixPy is based on the Python language, which runs on a Linux system. The camera, images, models, and applications all require a large amount of memory. Since memory is limited, understanding memory usage and management methods is very important.

We can obtain the current memory usage status in many ways, either using built-in MaixPy methods or common Linux methods. For example, using Python:

from maix import sys
print(sys.memory_info())

The output is similar to:

{'cma_total': 0, 'cma_used': 0, 'cmm_total': 2147483648, 'cmm_used': 177512448, 'hw_total': 4294967296, 'total': 2060726272, 'used': 339562496}

Or

import psutil

# Get virtual memory info
mem = psutil.virtual_memory()

print(f"Total memory: {mem.total / (1024 ** 3):.2f} GB")
print(f"Used memory: {mem.used / (1024 ** 3):.2f} GB")
print(f"Available memory: {mem.available / (1024 ** 3):.2f} GB")
print(f"Memory usage: {mem.percent}%")

You can also use command line tools like cat /proc/meminfo or the free command to see memory info.

From the total and used fields, you can see the total available memory and the memory already used.

Note that the memory shown here is the memory available to Linux user space, which is less than the actual physical memory. For example, for a MaixCAM2 device with 4GiB memory, the default shown here is 2GiB. Why this is so will be explained below.

MaixPy MaixCAM Memory Layout

Because we use a Linux system, memory is divided into several regions based on usage:

The two types of memory to focus on are:

• Linux User-Space Memory: This is the total memory available for our code and applications, such as allocating arrays, creating objects, and loading programs.

• CMM/CMA Memory: For the MaixCAM series, without a GPU, CMA is usually zero (not used). Chip manufacturers tend to define their own standards. For example, MaixCAM and MaixCAM2 both use a custom memory region for camera, NPU, and other hardware drivers that need frequent large memory access. This improves performance and reduces memory fragmentation. For example:

  • Camera frame capture: The image is first read into this memory region; if the image needs to be viewed in Linux user space, it is mapped or copied into user space. Thus, the higher the camera resolution set by the application, the more memory is required.
  • NPU running models: Models are loaded into this custom memory area, not user space. For instance, if a large LLM model requires 1.5GiB, this memory area must be at least 1.5GiB to load the model successfully.

Default Memory Sizes for Different Devices

Adjusting Memory Allocation

As mentioned above, default values are generally sufficient for most applications. If you want to adjust sizes, for example to increase CMM memory for loading larger models, you can modify the allocation yourself.

Because CMM is typically designed by the CPU manufacturer, the modification methods differ by device:

• MaixCAM: The CMM memory is actually called ION memory by the vendor. Modifying it is complex and requires recompiling the system. See GitHub modification reference.
• MaixCAM2: The CMM memory is called CMM memory by the vendor. The MaixCAM2 image has been optimized so you only need to modify the maix_memory_cmm value in /boot/configs to the desired size in MiB. The default value is -1, which means using the default allocation.

If you need me to translate the content into a more formal technical document style or adapt it for a specific audience, let me know!