Pushing performance

Exaggerated claims for fast processors seem a thing of the past, but new realities may continue to limit their use.

Th 0601vsd Andywilson

Exaggerated claims for fast processors seem a thing of the past, but new realities may continue to limit their use.

Twenty years ago, the machine-vision market was very different. Board-level vendors, intent on offering their customers the fastest processors on the market, were only too willing to announce products that incorporated the latest processors fresh from semiconductor factories. In trying to gain market share, some of these board vendors decided to promote products that often were less than working prototypes.

One company, I clearly remember, decided that the latest bit-slice processor from AMD (Sunnyvale, CA, USA; www.amd.com) was the processor of choice. In a massive series of advertisements, this multibus coprocessor was gradually “unveiled” to the readers in a fashion rather like the monolith from Stanley Kubrick’s movie2001: A Space Odyssey.

Although the series of ads touted the powerful features of the board, the truth was that the F-register on the processor was stuck. It could not work without a hardware register patch. Despite this, marketing managers were duly deployed to trade shows to present demos of how the board worked. Unfortunately, when questioned about the processor itself, all these folks could promise was that a revision of the processor would be available “within weeks.” Weeks turned to months, months turned into a year, and finally the product worked-but only if you knew how to program a bit-slice engine!

Back then, few software or development tools were available. But for those engineers intent on increasing system throughput, such boards did find their way into high-end military and medical products.

Today, the machine-vision and image-processing markets are very different. Given the emergence of high-performance gate arrays with embedded PCs, many sophisticated applications can be performed using a single FPGA. And nearly every frame-grabber and camera vendor uses these devices to tailor their products to meet a customer’s demands.

Some vendors are now starting to offer high-level graphical pipelined software tools to program these devices in an intuitive manner (see “Visual applets make FPGA programming a snap,” p. 48). Furthermore, a number of semiconductor companies have recently announced products that further leverage the power of reconfigurable computing (see “New processors push imaging performance,” p. 16). “Besides fine-grain reconfigurable devices such as FPGAs,” says Doug Pihl, CEO of MathStar (Minneapolis, MN, USA; www.mathstar.com), “these are generally either processor-centric clusters that use some kind of RISC core along with application-specific logic blocks-which may or may not be reconfigurable-or mid-grained reconfigurable products.”

However, machine-vision board vendors have been reluctant to adopt these products. With the advent of large FPGAs, board vendors have found that image-processing functions can be used to solve a large number of machine-vision applications.

In addition, despite offering software-development tools, semiconductor vendors often target large-volume applications. Consequently, the application notes and benchmarks they provide are frequently targeted far from the designer of image-processing and machine-vision systems.

Last, and perhaps most important, is the lengthy software-development time required to port existing machine-vision software libraries to new designs. With off-the-shelf CPU processing speeds and PowerPC processors now being embedded into FPGAs, it is far easier for board vendors to incrementally update existing high-speed image-processing boards than to introduce designs based on more radical technologies.

For military and medical applications that demand ultrahigh throughput of functions such as the FFT, however, these processors will find use. But they may only slowly emerge on board-level products designed for the OEM market. More likely, they will be used in embedded systems by companies that can afford both the integration and programming time required to build high-performance applications.

Th 0601vsd Andywilson
Click here to enlarge image

Andy Wilson

More in Life Sciences