From 10GigE to 200GigE: Scaling Industrial Vision Without Rebuilding the System

This article was originally published in inspect WORLD OF VISION. 

Industrial vision systems are being asked to process increasing amounts of information in less and less time. Production lines are becoming faster, sensors are delivering higher resolutions and frame rates, and AI and deep-learning methods are increasingly being used to detect defects and evaluate process behaviour directly during production.

Emergent Vision Technologies has been working on the data-transfer technology required for this for more than 16 years. In 2010, the company introduced the first 10GigE industrial cameras, followed by the first 25GigE cameras in 2018 and the first 100GigE cameras in 2020. At VISION 2026, that progression extends to 200GigE with the HYPERION area-scan and STRATOS line-scan platforms.

These developments represent more than successive increases in interface speed. Behind them are many years of experience building systems that must acquire, transfer and process large image streams reliably—particularly when multiple cameras are operating in parallel. Today’s 10GigE, 25GigE, 100GigE and 200GigE platforms are based on a common GigE Vision architecture. This allows performance to scale without introducing fundamentally different camera and networking technologies for each bandwidth class.

Two Data Paths for High-Speed GigE Vision

There is more than one suitable approach to transferring large volumes of image data efficiently.

Emergent continues to support its own zero-copy implementation of the GigE Vision Streaming Protocol (GVSP), developed and optimized over many years for high-speed GigE systems. In parallel, current systems support GigE Vision 3.0 with RDMA/RoCEv2.

Both approaches are intended to reduce unnecessary CPU load and memory operations. They also allow the system developer to select the data path that best suits the particular architecture. Operating system, network technology, camera count, existing infrastructure and the required processing chain can all create different requirements. An established GVSP system may have different constraints from a new system designed around RDMA from the outset. Emergent supports both approaches and can recommend which data path is better suited to a particular system architecture.

NVIDIA GPUDirect provides another option by allowing image data to be transferred directly into GPU memory for processing. Emergent supports GPUDirect on both Windows and Linux, including with its own network interface cards. This is particularly relevant when AI, image processing or reconstruction takes place on the GPU, since additional transfers through host memory can otherwise consume bandwidth and CPU resources.

The new 200GigE cameras illustrate how demanding these data-transfer requirements have become. The HYPERION AH-21000-G with the Gpixel GSPRINT4521 delivers 21 MP at up to 1,000 fps. STRATOS brings 200GigE to line-scan imaging: the LST-16KG5 provides 16K resolution, 256-stage TDI and line rates of up to 1 MHz.

For industrial inspection and R&D, this combination of spatial and temporal resolution can be relevant when the process becomes faster but small image details still need to be retained—for example in electronics and semiconductor inspection, surface analysis, materials testing or the study of fast mechanical processes.

Matching Bandwidth to the Inspection Task

The fastest available interface is not automatically the right choice for every machine.

For a compact inspection system, 10GigE may be sufficient. At higher process speeds or with larger sensors, 25GigE, 100GigE or 200GigE may be appropriate. Because the cameras operate within the same GigE Vision ecosystem, performance can be increased without switching to a completely different transport technology.

This is particularly clear with line-scan cameras. The compact EROS LE-4KG35 provides 4K line-scan imaging over 10GigE. At 25GigE, the ACCEL LB-8KG7 delivers 8K mono or colour imaging at up to 240 kHz. In the 100GigE range, the PINNACLE series reaches 16K at up to 400 kHz; the LZT-8KG5 reaches 1 MHz at 8K using TDI. STRATOS extends this progression to 200GigE.

This allows interface performance to be matched to line speed, resolution and inspection width. Web materials, electronics, wafers, printed products and other continuously moving materials can therefore be designed around the actual throughput requirement rather than being tied to a single interface class.

The same principle applies to area-scan cameras. New BOLT 25GigE and ZENITH 100GigE models using Sony Pregius S global-shutter sensors cover a broad performance range: from very high resolutions of up to 105 MP at 113 fps to variants exceeding 800 fps at around 12 MP. For large-area inspection or measurement tasks, resolution may be the priority, while sorting and fast process monitoring may benefit more from high frame rates.

Reducing the Programming Effort Around AI

AI and deep learning can extend the capabilities of industrial inspection. A trained model alone, however, does not make a production-ready vision system. Cameras must be controlled, buffers managed, image data distributed to the appropriate compute resources, GPUs coordinated, and results returned to the application.

Emergent eSDK Pro is intended to reduce this integration effort. FlexProc allows image-processing pipelines to be created in which processing tasks can be assigned to available CPU and GPU resources. FlexTrans manages the movement of image data between those processing resources—including across multiple GPUs and servers. Together, the two components provide a configurable framework for acquisition, processing and data transfer without requiring the user to program the entire low-level data path from scratch.

Technologies such as zero-copy transfer and NVIDIA GPUDirect can be used within this pipeline to reduce unnecessary memory operations and CPU load.

One system at VISION 2026 demonstrates this approach in concrete terms. Eight BOLT HB-25000-SB 25GigE cameras, each delivering 24.5 MP at up to 105 fps, connect via 25GigE and Emergent network interface cards to two servers. Each server contains two NVIDIA RTX 6000 Ada Generation GPUs, for a total of four GPUs.

eSDK Pro uses FlexProc to distribute segmentation and tracking tasks across these compute resources. FlexTrans manages image-data transport through the processing pipeline and between the individual system resources. Zero-copy transfer and GPUDirect reduce unnecessary data movement as images are transferred into GPU memory for processing.

The practical objective is to reduce the amount of infrastructure code the customer has to develop. A manufacturing or vision engineer should not have to spend most of a project implementing buffer management, assigning cameras to GPUs or handling data transfer between servers.

That development time can instead be spent on the part of the inspection task where the engineer’s domain knowledge matters most: Which deviation is actually a defect? Which product variation is acceptable? What behaviour on the production line indicates a process problem?

eSDK Pro replaces neither the AI model nor this process expertise. Instead, the software is intended to simplify the acquisition and processing infrastructure so that more development effort can be invested in training and validating the model.

A PCB inspection system demonstrates the same approach on a smaller scale. Four EROS 10GigE cameras inspect printed circuit boards. An anomaly-detection model identifies defective PCBs, while another software component controls a laser that marks the affected board.

In a pharmaceutical demonstration, two BOLT 25GigE cameras inspect blister packs in real time. An anomaly-detection model identifies missing or incorrect pills and displays the model confidence score for the respective inspection result.

When Compressed Images Are Suitable for Inspection

Increasing image-data volumes do not necessarily have to be handled solely through higher network bandwidth. In some applications, it is more useful to reduce the amount of data at the source. AEOLUS cameras combine sensor performance comparable to corresponding 25GigE cameras with integrated H.265 compression.

The AA-12000-SB-H with Sony IMX535 delivers 12.4 MP at up to 187 fps, while the AA-25000-SB-H with Sony IMX530 delivers 25 MP at up to 105 fps. Compression takes place directly inside the camera before the data stream is transferred to the host.

For inspection tasks, compressed images can be suitable when the relevant features remain clearly represented after encoding. Examples can include presence/absence checks, object classification, many anomaly-detection tasks, process monitoring or applications with long recording times.

The important point is that the inspection algorithm or AI model should be validated using the same compression settings that will later be used in production. Compression can alter fine textures, low-contrast details and individual pixel values. Its effect on the inspection result should therefore be evaluated for the specific application rather than assumed.

Uncompressed acquisition remains preferable where exact pixel values are required, very subtle or low-contrast defects must be detected, precision measurement is involved, or original image data must be retained for traceability or later re-analysis. Compression ratio, latency, image quality, model sensitivity and archival requirements should therefore be considered during system design.

AEOLUS therefore complements the high-bandwidth platforms rather than replacing them. HYPERION and STRATOS address applications in which large volumes of raw image data need to be transported. AEOLUS is intended for tasks where reducing the data volume is the more efficient system solution.

Experience Matters More as Bandwidth Increases

In the 16 years since Emergent brought the first 10GigE industrial cameras to market, available interface bandwidth has increased by a factor of twenty. The more difficult task, however, has remained fundamentally the same: making the camera, network interface card, memory architecture, GPU and software work together reliably.

That experience now includes both Emergent’s own zero-copy GVSP implementation and GigE Vision 3.0 with RDMA/RoCEv2, GPUDirect on Windows and Linux, multiple generations of Emergent network interface cards, and software for distributing acquisition and processing across multiple GPUs and servers.

For industrial inspection and R&D, this means that 10GigE, 25GigE, 100GigE or 200GigE can be selected according to the actual application while retaining a common GigE Vision architecture and a scalable data path.

Emergent Vision Technologies will demonstrate these systems at VISION 2026, 6–8 October in Stuttgart, Hall 8, Stand C36. See our show agenda.