CUDA ISP API performance

Follow us on:

Share This Page

Facebook Twitter LinkedIn WhatsApp Email Copy Link

Close

Send Email Using

Gmail (Web) Thunderbird (default mail app)

Close

CUDA ISP for NVIDIA Jetson
CUDA ISP for NVIDIA Jetson Basics
Getting Started
Evaluating the CUDA ISP Getting the code Building CUDA ISP for NVIDIA Jetson Building custom v4l2src element
User Manual
GStreamer
rrcudadebayer rrcudashift rrcudaawb
Examples
API usage GStreamer usage
Performance
Library GStreamer Elements
Contact Us

Library API performance

To measure the CUDA ISP API performance, we built a simple example (provided upon request) that iterates over the Apply methods for each algorithm and records performance metrics for each iteration. We measured the duration of each algorithm's Apply method. We also measured CPU, CPU RAM, GPU, and GPU RAM usage for the complete processing pipeline iterating at 30fps. We ran the experiments on both 1080p and 4K buffers. We also ran the experiments on the Jetson Orin Nano, Jetson Orin NX, Jetson Thor T5000, and Jetson AGX Orin.

Software performance measurement tools

• We measured the duration of each Apply method separately using the chrono library.
• We mainly used sys/times.h library to obtain the CPU usage. However, we used the proc/status file to obtain a secondary verification measure.
• We read the /proc/self/status file to obtain the CPU RAM usage.
• We used tegrastats to obtain the GPU usage.
• We used cudaMemGetInfo from CUDA to measure GPU RAM usage.

Every measurement is averaged over 100 iterations. The iterations are timed to run at 30 iterations per second.

Hardware setup

Platform	JetPack Version	Power Mode	GPU Frequencies	CUDA Version
Thor T5000	7.2	MAXN Performance (NVP Model 0)	1575000000	13.0
AGX Orin 32G	6.2	MAXN Performance (NVP model 3)	1300500000	12.6
Orin Nx 16GB	6.2	MAXN Performance (NVP model 0)	918000000	12.6
Orin Nano 8GB	6.2	MAXN_SUPER Mode (NVP model 2)	1020000000	12.6

For each system, we also used jetson_clocks to maximise the device clock frequency and thus the performance.

Results

The following table summarises CUDA ISP's performance results.

Algorithm	Jetson Thor T5000		Jetson AGX Orin		Jetson Orin NX		Jetson Orin Nano
Buffer size	1080p	4K	1080p	4K	1080p	4K	1080p	4K
Duration (ms) Kernel Launch -> Synchronize
CudaShift	0.27	0.52	0.42	0.78	0.66	2.08	0.62	2.30
CudaDebayer	0.29	0.44	0.39	0.68	0.60	1.44	0.57	2.03
CudaWhiteBalancer (Gray World Algorithm)	0.34	0.70	0.46	0.78	0.67	1.41	0.63	2.06
CudaWhiteBalancer (Histogram Stretch Algorithm)	0.38	0.72	0.50	0.97	0.81	1.60	0.77	3.03
Framerate (fps)
CudaShift	3496	1908	2375	1282	1522	550	1607	434
CudaDebayer	3424	2247	2570	1481	1647	469	1754	491
CudaWhiteBalancer (Gray World Algorithm)	2967	1430	2188	1283	1503	451	1592	485
CudaWhiteBalancer (Histogram Stretch Algorithm)	2631	1388	1984	1029	1218	368	1300	330
CPU usage (%)
CudaShift	0.512	0.519	0.124	0.228	0.321	0.407	0.428	1.28
CudaDebayer	0.507	0.519	0.136	0.254	0.339	0.416	0.438	1.23
CudaWhiteBalancer (Gray World Algorithm)	0.507	0.512	0.140	0.182	0.224	0.634	0.290	0.76
CudaWhiteBalancer (Histogram Stretch Algorithm)	0.519	0.514	0.142	0.191	0.227	0.730	0.315	0.80
CPU RAM (MB)
CudaShift	111.8	111.8	42.1	42.1	42.8	42.1	133.5	211.6
CudaDebayer	96.7	96.8	30.6	30.7	31.2	35.3	117.8	183,8
CudaWhiteBalancer (Gray World Algorithm)	102.6	102.6	34.6	34.7	35.4	31.1	123.7	187,7
CudaWhiteBalancer (Histogram Stretch Algorithm)	111.1	111.1	53.3	41.4	42.1	42.6	130,5	206.3
GPU usage (%) **
CudaShift+CudaDebayer+CudaWhiteBalancer	1.48	6.72	1.15	4.07	2.73	10.38	2.50	9.51
GPU RAM (MB)
CudaShift	112.7*	142.7*	13.9	11.1	11.3	9.2	15.6	11.8
CudaDebayer	110.7*	134.7*	0.97	1.0	1.2	3.6	1.2	1.1
CudaWhiteBalancer (Gray World Algorithm)	118.7*	154.7*	3.1	4.0	3.5	1.4	9.6	3.8
CudaWhiteBalancer (Histogram Stretch Algorithm)	118.7*	154.7*	19.9	11.6	11.1	11.9	15.1	20

*: On the Jetson Thor T5000, RAM usage is consistently reported above 100 MB when GPU acceleration is used. GPU utilization on the Thor T5000 is collected using NVML, while on the other evaluated platforms it is obtained from the filesystem-based GPU load metrics.

**: Using Nsight Systems (nsys), GPU usage is calculated by adding the execution time of the CUDA GPU kernels and dividing the resulting total by the target period at 30 FPS framerate.