ximea
The cameras using the PCI Express Gen 3 interface excel through their extraordinary speed based on the data bandwidth of up to 64 Gbits.
The whole range of models offers DMA (Direct Memory Access) and data transfer to GPU is possible on Linux.
This additionally helps with extremely low, almost nonexistent latency.
A major benefit is the ability to easily combine these cameras with SoC (system on a chip) modules utilizing GPU, like the ones from NVIDIA.
Recently NVIDIA released new modules called Orin, the world’s most powerful AI edge computers.
This opened up new performance levels for using XIMEA high speed and resolution PCIe cameras in embedded vision applications.
XIMEA's technology partner MRTech tested two camera models together with the Jetson AGX Orin module.
Below is a description of a test configuration transmitting (stream) images from the camera to a computer with the results achieved.
Picture 1. Setup with CB120CG camera, AGX Orin and the receiving laptop
Transmitting side:
CB120CG-CM-X8G3 camera with 12 MP (4096x3072) resolution and 333 Fps full speed potential
or
MX245CG-SY-X4G3-FF-HDR with 24.5MP (5320 x 4600) resolution and optional Dual ADC - HDR feature
Attached to NVIDIA Jetson AGX Orin 32GB
Orin software stack allowed the implementation of an image processing pipeline streaming images from the camera to the computer.
Processing pipeline
Picture 2. Receiving laptop
Receiving side:
MSI Raider laptop with NVIDIA GeForce RTX 3070 Ti GPU card
Software:
Picture 3. How to measure G2G latency
Test results
During testing several camera modes were used.
Changes were checked with different image demosaicing algorithms from the fastest HQLI to the highest quality MG.
Image processing time was measured as well as Orin CPU/GPU load, Orin power consumption, and the Glass-to-Glass (G2G) latency.
Plus, the H.265 encoder bitrate was adjusted to handle the rapidly changing picture.
Picture 4. IFF SDK logfile
You can find the benchmark results of the various below settings in this PDF attachment.
Picture 5. jtop Linux application to see Orin stats
CB120CG-CM-X8G3 camera and full frame (4096x3072) image at 80 FPS
• Input image: RAW8
• Demosaicing: fastest HQLI demosaic algorithm with a 5x5 processing window
• Image processing latency on Orin: 26 ms
• Average Orin CPU load: 6%
• Average Orin GPU load: 34%
• Orin power: 29W
• Image processing latency on the receiving side: 11 ms
• G2G latency: ~ 60 ms
and
• Input image: RAW16
• Demosaicing: high-quality MG demosaic algorithm with a 23x23 processing window
• Image processing latency on Orin: 33 ms
• Average Orin CPU load: 7%
• Average Orin GPU load: 72%
• Orin power: 45W
• Image processing latency on the receiving side: 11 ms
• G2G latency: 60-70 ms
CB120CG-CM-X8G3 camera and 4K UHD (3840x2160) image at 100 FPS
• Input image: RAW8
• Demosaicing: HQLI (5x5 window)
• Image processing latency on Orin: 18 ms
• Average Orin CPU load: 5%
• Average Orin GPU load: 20%
• Orin power: 26W
• Image processing latency on the receiving side: 9 ms
• G2G latency: 40-50 ms
and
• Input image: RAW16
• Demosaicing: MG (23x23 window)
• Image processing latency on Orin: 21 ms
• Average Orin CPU load: 7%
• Average Orin GPU load: 57%
• Orin power: 41W
• Image processing latency on the receiving side: 9 ms
• G2G latency: 50-60 ms
CB120-CM-X8G3 camera and Full HD (1920x1080) image at 250 FPS
• Input image: RAW8
• Demosaicing: HQLI (5x5 window)
• Image processing latency on Orin: 6.4 ms
• Average Orin CPU load: 5%
• Average Orin GPU load: 30%
• Orin power: 23W
• Image processing latency on the receiving side: 3 ms
• G2G latency: 20-30 ms
and
• Input image: RAW16
• Demosaicing: MG (23x23 window)
• Image processing latency on Orin: 7.4 ms
• Average Orin CPU load: 7%
• Average Orin GPU load: 45%
• Orin power: 33W
• Image processing latency on the receiving side: 3 ms
• G2G latency: 20-30 ms
MX245CG-SY-X4G3-FF-HDR camera and full 4K (4096x4096) image at 60 FPS
• Input image: RAW8
• Demosaicing: HQLI (5x5 window)
• Image processing latency on Orin: 32 ms
• Average Orin CPU load: 4%
• Average Orin GPU load: 32%
• Orin power: 29W
• Image processing latency on the receiving side: 12 ms
• G2G latency: ~ 100 ms
and
• Input image: RAW16
• Demosaicing: MG (23x23 window)
• Image processing latency on Orin: 40 ms
• Average Orin CPU load: 7%
• Average Orin GPU load: 76%
• Orin power: 45W
• Image processing latency on the receiving side: 12 ms
• G2G latency: ~100 ms
MX245CG-SY-X4G3-FF-HDR camera and 4K UHD (3840x2160) image at 120 FPS
• Input image: RAW8
• Demosaicing: HQLI (5x5 window)
• Image processing latency on Orin: 17.3 ms
• Average Orin CPU load: 5%
• Average Orin GPU load: 30%
• Orin power: 29W
• Image processing latency on the receiving side: 9 ms
• G2G latency: 40-50 ms
and
• Input image: RAW16
• Demosaicing: MG (23x23 window)
• Image processing latency on Orin: 21.2 ms
• Average Orin CPU load: 5%
• Average Orin GPU load: 70%
• Orin power: 46W
• Image processing latency on the receiving side: 9 ms
• G2G latency: 50-60 ms
Picture 6. XIMEA MX245CG-SY-X4G3-FF-HDR camera
MX245CG-SY-X4G3-FF-HDR camera and full 4K (4096x4096) image in Dual ADC HDR mode at 42.6 FPS
Note that 42.6 FPS is the maximum framerate for Dual ADC Non-Combined camera mode
• Input image: RAW16
• Demosaicing: MG (23x23 window)
• Image processing latency on Orin: 45.5 ms
• Average Orin CPU load: 38%
• Average Orin GPU load: 60%
• Orin power: 44W
• Image processing latency on the receiving side: 12 ms
• G2G latency: ~110 ms
Don't hesitate to get in touch with us if you need details of system configuration and test results.
The tests were carried out by MRTech.