Gpixel GMAX3265 Image Sensor Processing
Gpixel is a turn key supplier of advanced CMOS image sensors, developed by a multi-disciplinary team of image sensor experts in CMOS image technology. From the offices in Changchun, China (headquarters) and Antwerp, Belgium is Gpixel specialized in providing high-end CMOS image sensor solutions for industrial, professional, medical and scientific applications.
Founded in 2012 by experienced CMOS image sensor designers and semiconductor physicists, Gpixel company is committed to continuously innovate and work in close cooperation with its customers and business partners to deliver the most state-of-art CMOS image sensor technologies and products to the global market.
One of the latest releases of Gpixel is GMAX3265 image sensor. This is high resolution (65 MPix) image sensor featured with the latest low noise 3.2 µm charge domain global shutter pixel architecture.
Fig.1. Gpixel GMAX3265 image sensor
GMAX3265 is defined in close collaboration with leading industry partners in inspection vision systems, as such are we confident that the unique features of GMAX3265 will meet the most demanding requirements for industrial camera applications.
GMAX3265 offers 9344 × 7000 resolution, global shutter with ultra-low read noise of 2e-, more than 70 dB dynamic range, and very small dark current at room temperature. Thanks to the light pipe technology, the sensor exhibits excellent shutter efficiency of 1/30,000 and large angular response.
GMAX3265 image sensor solution is offered in high speed version and normal speed version. High speed version delivers 71 fps at 10-bit output, normal speed version delivers 31 fps at 12-bit output. GMAX3265 is designed with on-chip sequencer, supporting various exposure and Region of Interest (ROI) modes, tailored to inspection imaging needs, enabling easy and cost-effective integration for camera manufactures. GMAX3265 is assembled with 239-pin micro-PGA ceramic package for reliability and good heat dissipation and a double-sided ARC D263 glass lid.
The image sensor has 56 pairs sub-LVDS data output channel, each running at maximum 1.08 Gbit/s. GMAX3265 image sensor is capable to offer 71 fps at 10-bit output or 53 fps at 12-bit output. High resolution and fast frame rate lead to significant increase of system throughput various camera applications.
XIMEA CB654 camera with color GMAX3265 high speed image sensor from Gpixel
This is compact >8K camera for industrial and scientific applications:
Fig.2. XIMEA high resolution color camera CB654 with GMAX3265 image sensor from Gpixel
XIMEA CB654 Camera Specification
That camera is based on 65 megapixel image sensor GMAX3265 from Gpixel, has optical format 2.3" and global shutter. This is both high resolution and high speed CMOS camera with PCIe interface (no grabber is needed) which is working at 64 Gbps bandwidth. The camera has high dynamic range, low noise, minimal latency. Data transfer is implemented via DMA.
How to process frames from XIMEA CB654 camera?
That camera generates very high data rate. Maximum performance for the camera could be evaluated as 71 fps * 65 MPix = 4.6 GPix/s. This is very complicated task just to send that data stream from camera to host PC. PCI-Express Gen 3 x8 is the only choice to cope with such a stream. That's what XIMEA has brilliantly implemented and the task has been solved. By XIMEA API we can send all frames to PC in realtime.
If we need to get maximum data rate from that high resolution image sensor, the task becomes really complicated. Raw data stream of 4.6 GPix/s (which is equal to 5.8 GB/s, because we have 10-bit pixels) after demosaicing is tripled, so at the end of image processing pipeline we will have almost 14 GB/s stream in 24-bit RGB format! As soon as we do computations at 16 bits per color channel, we should process two times more. There are no chances to solve that task on CPU, though it is not easy to get a solution on GPU either, thoigh this is possible.
Fig.3. Fastvideo SDK to process raw images from XIMEA mono and color cameras with Gpixel image sensors
In most cases there is no need in realtime image processing for such applications, though the question about the performance is the key issue, especially for Gpixel GMAX3265 image sensor. We do need to get it as high as possible. If we consider for comparison, how Fastvideo SDK on good GPU can process frames from 48 MPix CMOSIS camera, we can get rough understanding about expected performance and utilized image processing pipeline.
The performance is strongly depends on GPU model and on complexity of the pipeline. In general, we can get the performance around 2-4 GPix/s per GPU and that gives us an idea of how many GPUs we will need for processing. Multi-GPU solution with Fastvideo SDK is the right choice for that task.
User can get trial version of Fastvideo SDK to check image quality and performance for sample images to evaluate processing time for a chosen image processing pipeline.
To create end-user software for cameras with high performance Gpixel image sensors, one could have a look at gpu-camera-sample project on Github. User can download source codes and add camera controls to build high performance application with GPU image processing. One can also download binaries to work with a particular camera or with raw images in PGM format from SSD. That software is based on Fastvideo SDK and it could work with all NVIDIA GPUs, including Jetson.
Benchmarks on NVIDIA GeForce RTX 2080TI
We've done some tests with the software gpu-camera-sample for raw frames from Gpixel GMAX 3265 image sensor. The pipeline included data acquisition, dark frame, FFC, linearization, BPC, white balance, L7 debayer, gamma sRGB, 16/8-bit transform, JPEG encoding (subsampling 4:2:0, jpg quality 90), viewport texture copy, monitor output at 30 fps. We've got total processing time on GeForce RTX 2080TI around 11 ms per frame which is more than 4 GPix/s. The same pipeline for 12-bit raw image could give us 13 ms processing time. These are preliminary results and we are working on optimization at the moment.
JPEG compression for that frame (image resolution 9433 × 7000, 24-bit, subsampling 4:2:0, jpeg quality 90) was done within 3.3 ms which corresponds to performance ~60 GB/s. This is the time for GPU processing without data import/export.