Computers for machine vision
In this section we will cover some of the system features that are important when specifying a PC and a supplier for a vision system.
Only a few years ago compact and embedded PCs just did not have the performance needed for a machine vision system. This meant that PCs for machine vision often standardised on the 19" rack mount approach with server grade high-performance components installed.
In recent years we have seen significant reductions in the power to performance ratio. Very capable fanless and compact industrial PCs without internal cables or internal moving parts can deliver per formance with high environmental specifications. They are suitable for both the harsh environments of the factory floor and the processing power required for imaging. As in most markets there is a wide variety of options and invariably you get what you pay for. Time and time again it proves to be a false economy specifying office grade PCs in an industrial machine vision system.
Recently, there has been a significant growth of non-Intel/Windows platforms in embedded applications such as networking, automotive and mobile phones. These ARM processors are now migrating into the machine vision market in high volume, low power applications. We are seeing machine vision cameras supporting the embedded MIPI interface and machine vision software vendors porting to these platforms. STEMMER IMAGING is embracing this change and providing support services and software modules to significantly reduce the time to market
Selecting the right components can make a huge difference to the performance of a vision system and its long-term reliability, especially as a vision system is expected to handle very large amounts of data (images) in quick succession. Not all PC systems are well suited for these tasks and most off-the-shelf domestic computers often have never been stress tested with continuous high throughput of data. Leaving the interface architecture to one side, as this is covered in the acquisition section in detail, some of the key considerations when specifying vision computing hardware are listed below.
- What is the maximum data rate that the system needs to handle?
- Number of cameras / image size / frame rate?
- Memory bandwidth required?
- How many simultaneous images need to be held in memory for
- Memory required?
- Does the system have to process this data at the same rate? *Combination of speed / number of processors / memory bandwidth?
- Does the system have to record all this data to disk?
- RAID: disk configuration, data rate?
- What level of hardware redundancy is required to ensure data integrity?
- RAID: disk mirroring, redundant power supplies?
- Does the software support multi-processor or multi-core systems?
- Number of processors and different configurations?
- What level of reliability MTBF (Mean Time Before Failure) has to be
- Consider the PC design including cables used or solid state disks.
- What is the environment of the target location in terms of temperature, vibration?
- Expected production start / life cycle?
- Component obsolescence and supply security?
- Mechanical considerations?
- Housing size and design?
- Environmental aspects?
- Lead-free, PC fans/filters, electrical noise, EMC, etc.
Selection of housing
Depending on the environment, a computer has to meet certain demands. When a system is deployed in a harsh industrial environment, considerations such as shock, vibration, thermal cycling, humidity and dust have to be considered, whereas a system used in a clean room can have lesser demands. Some environments specify extreme vibration resistance (for example mobile applications).
Other imaging applications demand stability even at extreme temperatures (for example traffic surveillance where the computers are installed outside). The same requirements have to be considered for the insides of a vision system.
There are many PC processor configurations, all with varying performances. In contrary to other applications, processor speed is not the only important factor in choosing a proper processor for an imaging application. In general suitable processors can be divided into three application areas:
- Workstation: Intel Core i7/i5/i9
- Mobile: Intel Atom & ULV Core i5 & i7
- Server: Intel Xeon
Workstation processors are most commonly found in standard imaging applications, while mobile processors are often to be found in embedded vision systems due to their low power consumption. For the most demanding applications that need multiple processors or higher data throughput rates, server processors are often used as they offer more memory ports, increasing the data flow through the processor.
With Moore's law doubling performance of processors every two years, applications which needed server based performance 5 years ago can now be served using a fanless compact PC with mobile low power processors. With this high level of innovation in PC chipsets we see relatively short lifecycles of the processors used in general computing. This is at odds with the demands of an industrial system that requires a long period of availability and long-term support. Companies such as Intel pick a few specific processors and chipsets to be part of their embedded product lines. These have extended availability at a slightly higher cost.
Number of cores and processors
In most cases a single chip includes a number of cores that allow different applications or tasks to run on each core.
In theory, this multiplies the performance of the PC compared to a single core based machine. This however is only beneficial for a vision application if the software supports 'multi-threading'. Multi-core processors offer a viable alternative to using two separate server processors such as the Intel Xeon. Typically, server motherboard designs support two to four processors, providing systems with 40 cores and more. The core count continues to increase as does memory bandwidth. Processors with a large number of cores are already available and we expect to see more in the future.
Typical performance comparison
With the range of new processors and new architectures available on the market, there is a lot to consider when selecting the right processor. The following diagram should provide a quick overview of the relative performance of different processors. Attention should be paid to the fact that different algorithms put different loads on the processors.
These CPU benchmark values use the PassMark® ratings. As with any benchmark system, it cannot tell the whole story, as many factors can potentially influence the final result, but it provides a good reference point to compare different processors. To indicate the growth in performance, the processing capability of a mid range desktop processor from 2012 can now be equalled by an entry level embedded ATOM processor with low power requirements, enabling fanless designs.