Research Companies and Industries Find Business People Take The Tour Subscription Options FPGAs are the power behind image and video processing
04/11/2006 12:04:09 PM EDT
Electronics Weekly
The Institute for Electronics, Communications and Information Technology (ECIT) at Queen’s University, Belfast combines the capabilities of research groups including system-on-chip; speech, image and vision; and high performance computing. ECIT houses more than 120 researchers and, with around £38m investment, has major implications for the economic development of Northern Ireland.
Approximately 25 of ECIT’s researchers are involved in development work related to FPGAs – highly specialised devices that present significant advantages over traditional high performance computing (HPC) platforms. The researchers are focusing in particular on the use of FPGAs in high performance image and video processing, real-time video processing, new ways of developing silicon IP cores, and the application of FPGAs to programmable networks.
“For image and video processing we have several applications that need really high power computing. For instance, in processing ultra-large medical images, such as the very high resolution scans used to identify cervical cancer, some of the images can be 100k x 100k pixels, which requires a vast amount of processing power,” explains Professor Danny Crookes, director of the Speech, Image and Vision Systems group.
“Normally these images are 2D, but because the samples themselves can be 3D, what you sometimes find is that different parts of the image are either in or out of focus. We’re therefore aiming to get 3D information by using computational techniques to reduce the number of scans needed at different heights.”
Crookes says that in real-time video processing they are looking at applications involving multiple video sources being processed simultaneously.
“For example, for security surveillance where you might have multiple cameras and want to model the movements of people to recognise potentially suspicious behaviour,” explains Crookes. “We’re also investigating the use of mobile camera devices with local on-board processing for tracking and recognition – where power, portability and speed are obviously crucially important.”
Why FPGAs?
In the past, CrookesÂ’ team used supercomputer clusters to give them the performance they needed, and research concentrated on developing software tools - particularly programming languages and libraries - to enable applications to be written at a high level and then compiled onto a parallel system. In recent years, however, the teamÂ’s focus for HPC has turned to the use of FPGAs.
“The advantages of FPGAs over a traditional supercomputer really depend on the type of processing you’re doing, but for many applications, particularly image and video processing, you can get the same performance with an FPGA that you would with a supercomputer cluster – and with the added benefits of portability, lower cost, and low power if you want mobile devices,” continues Crookes. “The disadvantage, however, is that the design time for designing efficient hardware is much higher than for software, and the skill set required is relatively rare - particularly among application developers.”
Crookes says that if the best design approach is not used when designing for software then a factor of two or three can be lost. For hardware the loss could be a factor of 1,000.
“We’re therefore aiming to develop software tools providing a standard application development environment - for which we’re using Matlab as the front end, but with a sophisticated reasoning engine and a way of describing ‘intelligent’ design components and design patterns - which will result in a very efficient hardware implementation - almost automatically,” adds Crookes.
Why Silicon Graphics (SGI)?
By avoiding the previous limitations of FPGAs, SGIÂ’s RASC (reconfigurable application-specific computing) approach presents multiple benefits for the ECIT teams.
“Because RASC can access memory at really high speeds, in your design you can focus your attention on the problem you want to solve, rather than on how to get around the limitations of the hardware you’re using,” explains Crookes.
“The other big advantage is that because it’s scalable, you can have access to very large memory and don’t have to worry so much about things like synchronising access to different memory banks. This too means that you have one less thing to think about when you’re designing your system.”
The shared memory model is considered another bonus which means the HPC cluster’s view of memory can be the same as that of the RASC. “So we can communicate much more easily between the two,” says Crookes. “We can also move the processing from the cluster onto the RASC without physically changing the memory structure. If you couldn’t do this, you’d end up with two separate environments with different libraries.”
One of the issues faced with FPGAs can be the memory interface.
“What we’re finding is that one of the most significant design issues is the memory model. SGI have done a lot of forward thinking on this, and made it much more possible to get a standard memory model that saves a lot of development time,” states Crookes. “It’s already proving its worth in the HPC arena, and I think that for the same reason, for what we’d call ‘significant’ systems, we can expect it to more than double the speed of our total systems development.”
RASC also enables a ‘hybrid’ implementation where part of an application will run on an FPGA, and part on the high performance cluster, so maximising the use of resources. “The FPGA is very good at certain things, and we can use this to free up the rest of the cluster for doing things with the data for which the processors are better suited.”
Looking to the future, Crookes and his team will be using RASC with their SGI Altix server and Silicon Graphics Prism visual computing solutions to develop a variety of specialist applications and tools. “A lot of supercomputer vendors are now offering FPGA facilities and components, and we’re working in conjunction with a world leader in high performance scientific libraries to develop FPGA implementations of these same libraries,” says Crookes. “This will mean that the user, without having to change their code, can run it on a mixture of FPGA and HPC, or even switch seamlessly between the two.”
Rob Cowieson is account manager at Silicon Graphics
hoovers.com |
