Birds Eye View - Performance Measurements on NVIDIA Jetson AGX Thor, AGX Orin and Orin NX platforms

Follow us on:

Share This Page

Facebook Twitter LinkedIn WhatsApp Email Copy Link

Close

Send Email Using

Gmail (Web) Thunderbird (default mail app)

Close

⇦ Performance/Profiling

Home

Performance/NXP iMX8 ⇨

Getting Started
Quickstart Guide Evaluating the Product How to Get the Code
How BEV Works
Overview Background Birds Eye View and Libpanorama
User Guide
Building and Installation Minimal Application
Calibration Guide
Understanding the Calibration Process Preparing the Calibration Images The Calibration Tool GUI Performing the Calibration
Calibration Guide Legacy
GStreamer
GstBEV Plugin Generic Pipelines Jetson Pipelines iMX8 Pipelines
Performance
PC Profiling NVIDIA Jetson Thor and Orin NXP iMX8
Contact Us

RidgeRun's Birds Eye View (BEV) on NVIDIA Jetson platforms delivers real-time or near-real-time surround-view processing depending on the Jetson model, input resolution, number of cameras, and power configuration. In the benchmarks on Jetson AGX Thor, Jetson AGX Orin, and Orin NX, BEV handled 4-camera and 6-camera pipelines from 1024x768 up to 4K, reaching real-time performance for many HD and Full HD cases and approaching real time for selected 4K workloads. This page summarizes the benchmark environment, measurement method, and the observed latency, framerate, CPU usage, GPU usage, and RAM usage so developers can estimate how BEV may scale on Jetson-based embedded systems.

The last section of this page presents a comparison of the performance when running Birds Eye View on a Jetson Thor T5000, Jetson Thor T4000 and a Jetson AGX Orin, for 4 and 6 video inputs at 4k@30fps.

Benchmark Environment

Hardware:

• A NVIDIA Jetson AGX Thor, Jetson AGX Orin and a Jetson Orin NX.

Software:

• Jetpack SDK
• GStreamer

Methodology

• CPU and RAM usage measured using PID/STATUS
• Framerate measured using GstPerf
• GPU usage measured with nvidia-smi on Jetson Thor and tegrastats on Orin
• Maximum performance power mode enabled (all cores active) with

  sudo /usr/sbin/nvpmodel -m 0

  .
• Jetson clocks enabled/disabled with:

  sudo /usr/bin/jetson_clocks

  .
• The pipeline used for the tests is:

  GST_DEBUG=2 gst-launch-1.0 bev name=bev0 calibration-file=birds_eye_view.json \
   filesrc location=$INPUT_0 ! qtdemux ! h264parse ! nvv4l2decoder ! nvvidconv ! $CAPS ! bev0.sink_0 \
   filesrc location=$INPUT_1 ! qtdemux ! h264parse ! nvv4l2decoder ! nvvidconv ! $CAPS ! bev0.sink_1 \
   filesrc location=$INPUT_2 ! qtdemux ! h264parse ! nvv4l2decoder ! nvvidconv ! $CAPS ! bev0.sink_2 \
   filesrc location=$INPUT_3 ! qtdemux ! h264parse ! nvv4l2decoder ! nvvidconv ! $CAPS ! bev0.sink_3 \
  bev0. ! perf print-cpu-load=true  ! fakesink -v

Jetson AGX Thor Performance Results

This section presents the performance results obtained with the NVIDIA Jetson AGX Thor and RidgeRun's Birds Eye View.

Latency

Table 1 summarizes the latency added by the BEV element in the pipeline, for different resolutions and number of input video streams (4 and 6). The table includes the performance results when using or without using Jetson clocks.

Table 1: Latency of BEV running on Jetson AGX Thor for different input resolutions.

Input/Output resolution	Number of input videos	Latency (ms)
		With Jetson Clocks	Without Jetson Clocks
1280x720	4	3.071	5.074
	6	4.689	5.308
1920x1080	4	5.894	10.363
	6	7.981	10.693
4K	4	18.856	24.041
	6	28.134	37.231

CPU, GPU and RAM Usage

In the table 2, you can see the CPU, GPU and RAM usage with and without Jetson Clocks for T4000 and T5000 variants from the Thor family with cases of 4 and 6 input video sources and a resolution of 4K at 60fps.

Table 2: Jetson AGX Thor CPU, GPU and RAM Usage percentage for variants T4000 and T5000

Platform	Number of input videos	CPU Usage (%)	GPU Usage (%)	RAM Usage (%)
T5000	4	1.19%	17.44%	1.65%
	6	10.93%	31.29%	3.79%
T4000	4	6.32%	18.93%	4.94%
	6	12.22%	31.33%	6.30%

Jetson AGX Orin and Orin NX Performance Results

Latency

Table 3 summarizes the latency added by the BEV element in the pipeline, for different resolutions and number of input video streams (4 and 6). The table includes the performance results when using or without using Jetson clocks.

Table 3: Latency of BEV running on Jetson AGX Orin for different input resolutions.

Platform	Resolution	Number of input videos	Latency (ms)
			With Jetson Clocks	Without Jetson Clocks
Orin NX	1280x720	4	7.313831	10.511158
	1920x1080	4	9.759018	14.606874
	4K	4	33.144881	53.896635
AGX Orin	1280x720	6	9.77772	15.066253
	1920x1080	6	13.639131	20.51331
	4K	6	45.892468	65.539032

CPU Usage

In Table 4 you can see the performance with and without Jetson Clocks for different platforms from the Orin family with cases of 4 and 6 input video sources with a resolution of 1024x768 at 60fps.

Table 4: CPU Usage percentage for AGX Orin and Orin NX

Platform	Mode	Number of input videos	CPU Usage (%)
			Avg	1	2	3	4	5	6	7	8	9	10	11	12
Orin NX	Without Jetson Clocks	4	16%	15%	19%	11%	17%	-	-	-	-	-	-	-	-
		6	21%	19%	22%	21%	21%	-	-	-	-	-	-	-	-
	With Jetson Clocks	4	12%	11%	12%	12%	12%	-	-	-	-	-	-	-	-
		6	16%	15%	18%	16%	16%	-	-	-	-	-	-	-	-
AGX Orin	Without Jetson Clocks	4	6%	1%	29%	14%	0%	0%	0%	0%	0%	-	-	-	-
		6	5%	1%	7%	34%	0%	0%	0%	0%	0%	-	-	-	-
	With Jetson Clocks	4	5%	1%	5%	25%	8%	0%	0%	0%	0%	-	-	-	-
		6	5%	1%	15%	17%	4%	0%	0%	0%	0%	-	-	-	-

GPU and RAM Usage

In table 5, you can see the GPU and RAM usage with and without Jetson Clocks for different platforms from the Orin family with cases of 4 and 6 input video sources with a resolution of 1280x720 at 60fps.

Table 5: GPU and RAM Usage percentage for Jetson AGX Orin and Orin NX

Platform	Mode	Number of input videos	GPU Usage (%)	RAM Usage (%)
Orin NX	Without Jetson Clocks	4	70.29%	7.18%
		6	77.07%	7.75%
	With Jetson Clocks	4	76%	7.21%
		6	79%	7.82%
AGX Orin	Without Jetson Clocks	4	65%	4.46%
		6	65%	4.94%
	With Jetson Clocks	4	69%	4.51%
		6	69%	4.88%

Framerate

Figures 1-4 present the results for the average Frames per Second (fps) for 4 and 6 input videos of different resolutions with or without jetson_clocks running. For many cases, Birds Eye View is capable of real time processing.

1024x768

RidgeRun's Birds Eye View keeps a high framerate for 6 inputs at 1024x768 as shown in Fig. 1, where it can reach up to 97 fps with Jetson clocks and 82 fps without Jetson clocks.

Fig. 1: FPS on Jetson AGX Orin for 4-6 1024x768 video inputs with and without jetson_clocks.sh

1080x720

The framerate for 1080x720 is slightly higher than for 1024x768 as showin in Fig. 2, where for 6 input videos, Birds Eye View reaches up to 98 fps with Jetson clocks and 85 fps without it.

Fig. 2: FPS on Jetson AGX Orin for 4-6 1080x720 video inputs with and without jetson_clocks.sh

1920x1080

For full HD inputs, Birds Ey View is capable of 60 fps for 6 input cameras without Jetson clocks, and nearly 70 fps with Jetson clocks as shown in Fig. 3. If the inputs are reduced to 4 video streams, the framerate can reach up to 97 fps with Jetson clocks and 83 fps without it.

Fig.3: FPS on Jetson AGX Orin for 4-6 1920x1080 video inputs with and without jetson_clocks.sh

4K

In the case of 4K input videos, Birds Eye View can achieve nearly 30 fps for 4 camera inputs using jetson clocks, as shown in Fig. 4.

Fig. 4: FPS on Jetson AGX Orin for 4-6 4K video inputs with and without jetson_clocks.sh

NVIDIA Jetson Thor vs Jetson Orin - Birds Eye View Performance Comparison

In this section we present a comparison of the performance obtained when running RidgeRun's Birds Eye View on Jetson Thor T4000 and T5000 variants and Jetson AGX Orin. The results presented were obtained using 4 and 6 input video streams at 4K@30fps.

CPU Usage

Fig. 5 presents the CPU usage results. Thor T5000 achieves the lowest CPU usage with 10.93% when running with 6 input video streams and 1.19% when running with 4 input video streams. The CPU usage of the Thor platforms is more than 10% lower than the AGX Orin.

Fig. 5: CPU Usage on AGX Orin, Thor T4000 and Thor T5000 for 6(left) and 4(right) input video streams at 4K@30 fps

GPU Usage

Fig. 6 shows the GPU usage results, in this case Thor T5000 and T4000 have nearly identical GPU usage, with around 31% for 6 video inputs and around 18% for 4 video inputs. The GPU usage of the Thor platforms is around 10% lower than the AGX Orin.

Fig. 6: GPU Usage on AGX Orin, Thor T4000 and Thor T5000 for 6(left) and 4(right) input video streams at 4K@30 fps

RAM Usage

Fig. 7 presents the memory usage. Thor T5000 uses the least memory with nearly 4% for 6 video inputs. Jetson AGX Orin uses around 10%, and Thor T4000's usage falls between the other two platforms.

Fig. 7: RAM Usage on AGX Orin, Thor T4000 and Thor T5000 for 6(left) and 4(right) input video streams at 4K@30 fps

Maximum Processing Framerate

Fig. 8 shows the maximum framerate achievable by the Birds Eye View element with 4K inputs. For 6 inputs, Thor T5000 is capable of 60 fps, while Thor T4000 can achieve 44 fps and AGX Orin is capalbe of 34 fps.

Fig. 8: Maximum processing framerate on AGX Orin, Thor T4000 and Thor T5000 for 6(left) and 4(right) input video streams at 4K@30 fps

FAQ

What does this page measure for Birds Eye View on NVIDIA Jetson?

This page reports BEV performance on NVIDIA Jetson AGX Thor, Jetson AGX Orin, and Orin NX. It summarizes latency, framerate, CPU usage, GPU usage, and RAM usage for multi-camera surround-view pipelines using 4 or 6 input video streams across resolutions from 1024x768 to 4K.

Which Jetson platforms were benchmarked?

The benchmarks include NVIDIA Jetson AGX Thor, Jetson AGX Orin, and Jetson Orin NX. The Thor results also include T4000 and T5000 variants in the platform comparison section.

How many cameras were used in the tests?

The page includes test cases with 4 and 6 input cameras. Performance changes noticeably as the number of input streams increases, especially at higher resolutions such as 4K.

Can BEV run in real time on NVIDIA Jetson?

Yes, for many HD and Full HD scenarios BEV reaches real-time performance on Jetson platforms. The page also shows that selected 4K cases can approach real time, but the exact result depends on platform, number of inputs, Jetson clocks, power mode, and the complete pipeline configuration.

What is the effect of enabling jetson_clocks?

Enabling jetson_clocks generally improves latency and increases the maximum achievable processing framerate by locking the platform closer to its maximum performance state. The tables on this page show lower latency and higher throughput in many cases when Jetson clocks are enabled.

How was framerate measured?

Framerate was measured with [[1]]. CPU and RAM usage were measured using PID/STATUS, GPU usage was measured with nvidia-smi on Jetson Thor and tegrastats on Orin, and the tests were run with maximum performance power mode enabled.

What pipeline was used for the measurements?

The benchmarks used a GStreamer pipeline with the BEV element, H.264 file inputs, NVIDIA hardware decoding through nvv4l2decoder, color conversion with nvvidconv, and performance measurement through perf. This means the published results reflect both the BEV processing stage and the surrounding accelerated media pipeline.

What latency was observed on Jetson AGX Thor?

On Jetson AGX Thor, BEV latency ranged from about 3.071 ms for 1280x720 with 4 inputs and Jetson clocks enabled to about 37.231 ms for 4K with 6 inputs without Jetson clocks. This shows that latency scales with both resolution and camera count.

What latency was observed on Jetson AGX Orin and Orin NX?

On Orin NX and AGX Orin, the reported latency ranged from about 7.314 ms for 1280x720 with 4 inputs on Orin NX with Jetson clocks enabled to about 65.539 ms for 4K with 6 inputs on AGX Orin without Jetson clocks. Lower resolutions and enabled Jetson clocks produced the best latency results.

Which Jetson platform gives the best 4K BEV performance in this page?

In the comparison section, Jetson AGX Thor T5000 shows the strongest 4K results among the listed platforms. For 6 input streams at 4K@30 fps, it reaches up to 60 fps maximum processing rate, while Thor T4000 reaches 44 fps and AGX Orin reaches 34 fps.

How much CPU and GPU load does BEV add on Jetson?

The answer depends on platform, resolution, and number of inputs. For example, the page shows Thor T5000 at 4K@60 fps using as little as 1.19% CPU and 17.44% GPU with 4 inputs, while higher-input cases and Orin-family platforms can use substantially more GPU and CPU resources.

Does BEV performance depend more on resolution or on camera count?

Both matter, but this page shows that increasing resolution and increasing the number of input cameras both raise latency and resource use. The most demanding cases are the 4K and 6-camera configurations, which consistently show higher latency and lower maximum throughput.

Can I use these numbers as a guaranteed result for my product?

No. These benchmarks are a sizing reference for the tested platforms and pipeline configuration. Actual performance will vary with the Jetson SKU, JetPack version, GStreamer version, codec settings, memory bandwidth, number of cameras, calibration data, and any other processing stages in the full application.

Where should I go next if I want to tune or compare BEV performance?

For additional context, review Birds Eye View/Performance/Profiling, Birds Eye View/Performance/PC, and Birds Eye View/Performance/NXP iMX8. For pipeline integration details, see Birds Eye View/GStreamer/Jetson Pipelines. If you need help optimizing BEV for your Jetson design, visit Birds Eye View/Contact us.