Jetson AGX Orin RAM Consumption: JetPack 7.2 vs. a Minimal Yocto Image
- Carolina Trejos
- 3 hours ago
- 5 min read
Memory consumption is an important design consideration for edge AI systems. Every megabyte used by the operating system reduces the memory available for applications, video pipelines, AI models, and other workloads.
This report compares system RAM consumption on the NVIDIA Jetson AGX Orin using two software environments:
NVIDIA JetPack 7.2
A minimal Yocto-based image derived from core-image-minimal
The initial comparison covers two scenarios:
An idle system baseline with no test pipeline running
A simple GStreamer UDP pipeline
The results show that the minimal Yocto image provides a significantly lower idle memory footprint.
Key Results
During the baseline test, JetPack 7.2 consumed an average of 1,650.75 MB of system RAM, while the minimal Yocto image consumed 1,120.75 MB.
This represents:
An absolute reduction of 530 MB
A relative reduction of approximately 32.1%
During the GStreamer test, the Yocto system also maintained a lower absolute RAM footprint:
Metric | JetPack 7.2 | Yocto |
Idle system RAM consumption | 1,650.75 MB | 1,120.75 MB |
System RAM with GStreamer pipeline | 1,709.98 MB | 1,559.98 MB |
Test Environment
The tests were performed on the same NVIDIA Jetson AGX Orin platform.
The JetPack environment represents a complete NVIDIA development platform. The Yocto environment was based on a minimal image and extended primarily with GStreamer and the utilities required to collect memory measurements.
The same general GStreamer pipeline configuration was used in both environments.
Test Scenarios
System Baseline
The system remained idle with no test pipeline running.
This test establishes how much RAM is consumed by the operating system, background services, shared libraries, and platform components before the application workload begins.
Simple GStreamer UDP Pipeline
A simple GStreamer UDP pipeline was executed while both system-level and process-level memory statistics were collected.
System RAM consumption was calculated using:
MemTotal - MemAvailable
This provides a more useful estimate of system memory pressure than relying exclusively on the MemFree field, because Linux intentionally uses unused RAM for caches and reclaimable resources.
JetPack 7.2 Results
Baseline
Metric | Average | Minimum | Maximum | Variation |
System RAM consumed | 1,650.75 MB | 1,647.77 MB | 1,652.84 MB | 5.07 MB |
Test details:
Measurement duration: 122 seconds
Number of samples: 60
Sampling interval: approximately 2 seconds
Average RAM consumption: 1,650.75 MB
Percentage of total system memory: approximately 5.38%
The 5.07 MB difference between the minimum and maximum measurements indicates that the JetPack baseline remained stable during the test window.
Simple GStreamer UDP Pipeline
Metric | Average | Minimum | Maximum | Variation |
System RAM consumed | 1,709.98 MB | 1,682.46 MB | 1,740.98 MB | 58.52 MB |
Test details:
Number of samples: 141
Average system RAM consumption: 1,709.98 MB
Baseline system RAM consumption: 1,650.75 MB
The average increase relative to the JetPack baseline was: 1,709.98 MB - 1,650.75 MB = 59.23 MB
The pipeline therefore increased total system RAM consumption by approximately 59.23 MB over the JetPack baseline.
Yocto Results
Baseline
Metric | Average | Minimum | Maximum | Variation |
System RAM consumed | 1,120.75 MB | 1,120.75 MB | 1,120.75 MB | 0 MB |
Test details:
Measurement duration: 120 seconds
Number of samples: 60
Sampling interval: approximately 2 seconds
Average RAM consumption: 1,120.75 MB
Compared with JetPack 7.2, the minimal Yocto image consumed 530 MB less RAM while idle:
1,650.75 MB - 1,120.75 MB = 530.00 MB
The relative reduction was approximately:
((1,650.75 - 1,120.75) / 1,650.75) × 100 = 32.1%
This lower baseline is one of the primary advantages of using a purpose-built Yocto image for resource-constrained deployments.
Simple GStreamer UDP Pipeline
Metric | Average | Minimum | Maximum | Variation |
System RAM consumed | 1,559.98 MB | 1,540.98 MB | 1,682.46 MB | 141.48 MB |
Test details:
Number of samples: 140
Average system RAM consumption: 1,559.98 MB
Baseline system RAM consumption: 1,120.75 MB
The calculated increase relative to the Yocto baseline was: 1,559.98 MB - 1,120.75 MB = 439.23 MB
The total system consumption while the pipeline was running remained below JetPack.
Understanding the Yocto Pipeline Increase
The larger increase over the Yocto baseline is one of the most important findings in the current data.
A minimal system may defer loading many libraries, drivers, caches, and multimedia components until an application requests them. Consequently, starting the first GStreamer pipeline can produce a larger visible change from the baseline.
JetPack may already have some related libraries, services, or platform components loaded before the test begins. Their memory is included in the higher JetPack baseline rather than appearing as pipeline-specific growth.
Other possible contributors include:
Differences in GStreamer builds and enabled plugins
Different multimedia backends
Lazy driver or firmware initialization
Shared-library loading
Buffer pool configuration
DMA buffer allocation
Page cache growth
Allocator behavior
Kernel memory associated with media or networking
Services activated when the pipeline starts
For this reason, the baseline subtraction should not be interpreted as the exact amount of memory privately owned by the GStreamer process.
Why the Operating Systems Have Different Baselines
Root Filesystem and Operating-System Scope
JetPack is designed to provide a complete NVIDIA development environment. A typical installation may include an Ubuntu desktop environment, graphical services, development libraries, multimedia components, platform utilities, and supporting background services.
The Yocto image used in this test was derived from core-image-minimal. It contained only the packages and image features selected at build time and did not include a desktop environment.
As a result, the Yocto system starts fewer processes and generally loads fewer libraries during idle operation.
Background Services
A full JetPack environment may run services that are not present in a minimal Yocto image, such as:
Desktop login and session services
Network management services
Package update services
Crash reporting
Additional logging infrastructure
Container-related services
NVIDIA platform utilities
The Yocto image starts only the services introduced by its selected recipes, dependencies, and image features.
Display and Multimedia Initialization
A JetPack installation may initialize display, graphics, camera, codec, and other platform-related components during boot, even when no test pipeline is active.
A headless Yocto image may avoid initializing some of these components until they are explicitly required.
This can reduce idle RAM consumption while shifting part of the initialization cost into the application startup phase.
Current Conclusions
The current measurements support two conclusions.
First, the minimal Yocto image provides a substantially lower idle memory footprint. Its baseline was 530 MB lower than JetPack 7.2, representing a reduction of approximately 32.1%.
Second, Yocto maintained a lower absolute system RAM footprint while the GStreamer pipeline was running. The measured difference was approximately 150 MB.
The distinction between absolute and incremental consumption is essential:
Absolute consumption indicates how much RAM remains available to the complete deployed system.
Incremental consumption helps identify what becomes active when the application starts.
Process PSS helps estimate how much physical memory is proportionally attributable to the application processes.
For product deployment, the absolute footprint may be the most relevant measurement. For optimization and debugging, the incremental and process-level measurements are equally important.
Final Perspective
A minimal Yocto image can provide a meaningful reduction in operating-system overhead on the Jetson AGX Orin. In this test, it recovered approximately 530 MB of RAM at idle and retained a 150 MB advantage while a simple GStreamer UDP pipeline was active.
That additional memory can be valuable for larger AI models, additional video streams, deeper buffering, or other application features.
However, the results also demonstrate why system memory comparisons must look beyond a single average. Baseline consumption, steady-state consumption, process PSS, kernel allocations, and post-pipeline recovery all provide different parts of the complete picture.
The current results are promising for Yocto, but broader testing is needed before drawing conclusions about the memory efficiency of individual multimedia workloads.
Optimize your NVIDIA Jetson platform with a purpose-built Yocto image. Contact RidgeRun to discuss your deployment requirements.
