To optimize GPU and memory performance on a Jetson running ROS 2, first measure the complete robot workload, identify its limiting resource, and then change one factor at a time. A higher GPU clock or fewer message copies may help a particular pipeline, but neither guarantees lower end-to-end latency: memory bandwidth, CPU scheduling, I/O, power limits, and sustained thermals can all be the real constraint.
Establish exactly what you are measuring
Jetson performance settings and software compatibility depend on the module, board, and installed software. Before tuning, record the system and freeze the workload conditions so comparisons mean something.
As an Amazon Associate I earn from qualifying purchases.
- Jetson module and SKU, carrier board, and power supply.
- JetPack and Jetson Linux release, ROS 2 distribution, and ROS middleware implementation (RMW).
- Application build, sensor types, image or point-cloud dimensions, rates, and relevant queue settings.
- For inference workloads, the model and precision in use.
- Selected power mode, cooling and enclosure arrangement, and ambient conditions.
Use documentation for the installed release and exact device rather than copying a power-mode table or command sequence from another Jetson. NVIDIA’s Jetson software documentation index lists Jetson Linux 39.2.1 alongside versioned guides for earlier releases; the installed version determines which guide and options apply.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Choose application-level measures first
Measure sensor-to-result latency, throughput, and missed deadlines or dropped data over a representative period, including operation after the system has warmed up. Record memory use, temperatures, power, and CPU, GPU, and EMC (memory-controller) clocks or utilization where available. NVIDIA documents tegrastats and jetson_clocks --show for inspecting platform state; its performance guidance also calls for monitoring CPU, GPU, and EMC frequencies during stress testing.
#1 Best Overall
- The NVIDIA Jetson AGX Orin 64GB Developer Kit makes it easy to get started with Jetson Orin. Compact size, lots of connectors, and up to 275 TOPS of AI performance make this developer kit perfect for prototyping advanced AI-powered robots and other autonomous machines.
- The developer kit includes a Jetson AGX Orin 64GB module, and can emulate all the Jetson Orin modules. It supports multiple concurrent AI application pipelines with the NVIDIA Ampere GPU architecture, next-generation deep learning and vision accelerators, high-speed IO and fast memory bandwidth. Now you can develop solutions using your largest and most complex AI models to solve problems such as natural language understanding, 3D perception, and multi-sensor fusion.
- Jetson runs the NVIDIA AI software stack, and use-case specific application frameworks are available, including Isaac for robotics, DeepStream for vision AI, and Riva for conversational AI. You can save significant time with NVIDIA Omniverse Replicator for synthetic data generation (SDG), and by using NVIDIA TAO toolkit to fine-tune pretrained AI models from the NGC catalog.
- Jetson ecosystem partners offer additional AI and system software, developer tools, and custom software development. They can also help with cameras and other sensors, as well as carrier boards and design services for your product.
- With the computing capability of more than 8 Jetson AGX Xavier systems in a developer kit that integrates the latest NVIDIA GPU technology with the world’s most advanced deep learning software stack, you’ll have the flexibility to create tomorrow’s AI solution as well as today’s.
Keep the same workload, power mode, cooling, and measurement window for each comparison. A peak clock or a short burst of throughput does not establish that the robot will sustain that result.
Find the resource limiting the ROS 2 graph
GPU utilization alone cannot explain a slow pipeline. A workload may be limited by GPU computation, memory bandwidth, CPU scheduling, serialization or copying, thermal or power limits, or a sensor and I/O stage. These are diagnostic categories, not conclusions that can be drawn from one utilization reading.
| Possible constraint | What to examine | Useful controlled experiment |
|---|---|---|
| GPU computation | GPU activity and clocks alongside end-to-end latency and throughput. | Change one GPU workload or supported configuration at a time; check whether the complete graph improves, not just the GPU stage. |
| Memory bandwidth | EMC frequency and behavior alongside data volume, rates, and the timing of stages that move large messages. | Test a single change to image or point-cloud volume, conversion work, or message movement while preserving acceptable robot behavior. |
| CPU scheduling | CPU activity, clocks, and whether CPU-side stages or callbacks delay downstream work. | Vary one CPU-side workload or deployment choice and compare deadlines and end-to-end results. |
| Serialization, copying, or buffering | Process placement, message ownership, queue depths, rates, and retained message lifetimes. | Compare a suitable composed, intra-process graph with the existing layout, using the same input and measuring the whole pipeline. |
| Thermal or power limitation | Temperature, power, and CPU, GPU, and EMC clocks during sustained operation. | Compare supported power modes under the same warmed-up workload and cooling arrangement. |
| Sensor or I/O stage | Input timing and stage-by-stage delays alongside missed deadlines or drops. | Change one input rate or data-volume factor only if the robot’s freshness and loss requirements remain satisfied. |
For each experiment, alter one dimension, repeat the run, and compare the same application metrics. A clock increase that does not improve sustained sensor-to-result performance is not a useful optimization for that workload.
Recommended Free Tools
Tune power modes and clocks for sustained behavior
NVIDIA’s nvpmodel selects power modes supported by a particular device configuration; those modes can cap available resources. jetson_clocks can set static maximum CPU, GPU, and EMC clocks, show current settings, store them, and restore saved settings. Use these as controlled tuning and measurement tools, not as blanket fixes.
NVIDIA’s Orin guidance warns that MAXN does not prevent hardware throttling when total module power exceeds the thermal design budget. It therefore does not guarantee the best performance for every workload. EMC frequency behavior also responds to average bandwidth demand, driver requests, and thermal throttling, so GPU clock alone is not a complete account of memory performance.
Rank #2
- AGX Orin 64GB Development Kit makes it easy to get started with AGX Orin. Its compact size, rich interfaces, and AI performance of up to 275 TOPS make it ideal for building advanced AI robots and other autonomous machine prototypes.
- The development kit includes AGX Orin 64GB module and can emulate all Orin modules. It utilizes the Ampere GPU architecture, next-generation deep learning and vision accelerators, high-speed I/O, and fast memory bandwidth. You can leverage the largest and most complex AI models to develop solutions for problems such as natural language understanding, 3D perception, and multi-sensor fusion.
- Jetson runs AI software and provides application frameworks for specific use cases, such as Isaac for robotics, DeepStream for visual AI, and Riva for conversational AI. Using Omniverse Replicator for Synthetic Data Generation (SDG) can save you significant time; while fine-tuning pre-trained AI models from the NGC catalog using the TAO toolkit can further enhance your results.
- Yahboom offers four kits for users to choose from. The AIlarge model voice module utilizes examples of AI large models and multimodal models; it provides 1TB/2TB SSDs with pre-flashed driver image files; and an 8MP USB industrial camera for image processing.
- It offers various online and offline mainstream AI large model development materials. The system is pre-configured with AI vision examples, ROS case studies, and AI large models. It supports offline/online deployment of large models for voice interaction, real-time video analysis, and visual positioning, helping you quickly get started with localized AI agent development.
- Check the power modes and clock controls documented for the exact module and installed Jetson Linux release.
- Choose one supported mode or clock configuration to test; avoid changing several controls at once.
- Run the actual ROS 2 workload long enough to observe warmed-up behavior with the intended cooling and enclosure.
- Compare end-to-end latency, throughput, missed deadlines or drops, memory use, power, temperature, and clock stability against the baseline.
- Keep a configuration only if it improves the application’s sustained results within its power and thermal limits.
Commands that change power or clock settings may require elevated privileges and can affect the whole device. Follow the instructions for the installed release rather than assuming a command or mode applies to every Jetson.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Reduce avoidable ROS 2 copies and buffering
Test composition and intra-process communication where it fits
ROS 2 intra-process communication can avoid some message copies when compatible publishers and subscribers run in the same process. The ROS 2 project’s example demonstrates an eligible path using a std::unique_ptr publisher and subscriber and checking that message addresses match. The result depends on ownership and graph topology: multiple subscribers or other subscriber arrangements can require copies or change ownership behavior.
For high-bandwidth images or point clouds, test composition and enable intra-process communication for the relevant components, then verify the behavior using documentation for your ROS 2 distribution. Compare copy behavior and whole-graph performance rather than assuming that enabling the option eliminates all copies.
Keep queues and data volume aligned with robot requirements
Intra-process communication does not remove model memory, application buffers, middleware queues, or copies outside the eligible path. Inspect image dimensions, message rates, conversion stages, queue depths, and how long messages remain retained. Reduce data volume or queue capacity only when the resulting freshness, buffering, and loss behavior remain acceptable for the robot.
Process boundaries can still be the right choice when they support the deployment architecture or fault isolation. Compare communication layout alongside copy count, queueing, and the consequences of a component failure; avoiding a copy is not the only design objective.
Rank #3
- 【Core Parameters】★AI Perf:34-67 TOPS ★GPU:512-core NVIDIA Ampere architecture GPU with 16 Tensor Cores ★CPU:6-core Arm Corte-A78AE v8.2 64-bit CPU 1.5MB L2 + 4MB L3 ★Memory:4GB 64-bit LPDDR5 51 GB/s ★Storage: external NVMe via M.2 Key M (NOTE:SUB Board No SD Card Slot)
- 【Empowered by Large Al Model, Enhanced Human-Computer Interaction】Jetson Orin Super leverages three AI models and incorporates an AI voice interaction module. This multimodal visual system matches the scene being described, enabling environmental awareness and AI visual gameplay. Combined with a large-scale voice module and camera, it enables speech-to-text, semantic analysis, natural conversation, and real-time video analysis, enabling advanced embodied AI applications.
- 【AI Upgrade】Jetson Orin Nano series modules are compact in size but can deliver up to 34-67 TOPS of AI performance, with power consumption ranging from 7 watts to 25 watts. Compared to the Jetson Nano B01, it offers up to 80 times the performance and sets a new standard for entry-level edge AI.
- 【Highly compatible carrier board】Yahboom's carrier board is fully compatible with orin nano module. Compared to carrier boards that use Jetson Nano on the market, the newly upgraded circuit supports 25W power mode, which enables larger and more complex neural networks and fully leverages the performance of the core module. The resources, size, and interfaces of the Yahboom carrier board are consistent with the official board, with the only difference addition of power switch button.
- 【Tutorial materials provided】The JETSON system based on Ubuntu 22.04 provides a complete desktop Linux environment with accelerated graphics, supporting NVIDI-ACUDA 12.6, TensorRT 10.7.0, cuDNN 9.6.0, OpenCV 4.10.0, etc. The performance on AI LLM, VLM and visual Transformer is significantly improved compared with the previous generation.
Use supported acceleration tools, then profile the whole graph again
NVIDIA describes JetPack as the official Jetson software stack and lists CUDA, TensorRT, Nsight developer tools, and Isaac ROS among the relevant tools and technologies. NVIDIA characterizes Isaac ROS as hardware-accelerated ROS 2 packages for Jetson. They can be relevant to GPU-heavy vision, inference, and robotics workloads, but availability and installation support depend on the platform and software versions.
Check the documentation for the selected Jetson software release before choosing packages or following installation steps. After accelerating one stage, measure the complete ROS 2 graph again: the bottleneck may move to data transfer, CPU work, memory bandwidth, or another stage.
Compare candidates on the robot’s actual constraints
Judge each candidate configuration against the same baseline and workload. Use only power modes documented for the exact SKU, and record compatibility alongside performance so a faster but unsupported combination is not mistaken for a viable deployment.
- Sustained end-to-end latency and throughput.
- Missed deadlines and drops.
- Peak and steady memory use.
- Power draw.
- Thermal headroom and clock stability after warm-up.
- Compatibility with the Jetson module, JetPack or Jetson Linux release, and ROS 2 distribution.
- For communication-layout changes, process placement, copy behavior, queueing, and fault-isolation needs.
There is no broadly applicable performance multiplier for these steps: outcomes depend on the device, software versions, ROS 2 graph, workload, power configuration, and cooling. The useful result is a repeatable improvement measured on the system that will run the robot.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Free tools Windows power users keep installed
One-click scans. No signup required.
