Looking back at the past year, a striking trend emerges: Increasingly, the hardware used for signal processing is something other than a DSP. I made a quick survey of the year’s developments by skimming the archives on www.bdti.com/InsideDSP. I discovered that only about half of the chips BDTI wrote about in 2005 were DSPs. The rest of the chips were general-purpose processors, FPGAs, or other types of hardware. Admittedly, this hardly qualifies as a scientific survey, but the results reflect an undeniable reality.
Ironically, DSPs are facing growing challenges partly because of the growing importance of signal processing in a wide range of applications. Today, signal processing is found in everything from digital power supplies to portable video players. The prevalence of signal processing workloads has motivated processor vendors of all stripes to add DSP-oriented features to their processors. For example, the latest ARM and MIPS architectures have remarkable similarities to DSP architectures. As a result of this trend, DSPs are not clearly differentiated on the basis of processor architecture.
Processor architecture is also less of a differentiator because processor vendors’ solutions are more complex. Ten years ago, a typical solution consisted of a DSP processor and basic software tools. The main value of such a solution was the power of the DSP architecture. Today, a typical solution consists of a highly integrated SoC, sophisticated tools, operating systems, and more. With such a solution, the processor architecture is often hidden from the user under layers of hardware and software. Processor architecture is still important, but it has become only one of the many factors that determine a solution’s suitability for a particular application.
Signal processing applications are also increasingly complex. Ten years ago, signal processing applications typically incorporated very little non-signal-processing functionality. For these applications, a DSP was often the only sensible choice. Today, an application with signal processing workloads may also need to run Linux, handle 3D graphics, or perform packet processing. For applications with such diverse workloads, it is reasonable to consider a diverse range of architectures—and a DSP won’t always be the right solution.
For that matter, DSPs aren’t always the right choice even for applications that are narrowly focused on signal processing. Other approaches sometimes offer lower prices, better performance, or other advantages over DSPs on these applications. For example, FPGAs can beat DSPs in terms of cost and performance in many communications infrastructure applications.
Despite the challenges, DSPs are not about to go away. The fact that DSPs are purpose-built for signal processing workloads ensures that DSPs will continue to have important advantages for some of these workloads. For example, DSPs have clear advantages for many audio applications due to their appealing mix of performance, programmability, integration, and price.
For processor vendors, the message is clear: customers don’t care about fancy DSP architectures—they want solutions that solve their specific design problems. This means that processor vendors must clearly demonstrate how their solutions meet application needs.
And system designers must consider all relevant processors, regardless of what these processors are called. To do so, system designers must understand the strengths and weaknesses of many different kinds of architectures. And they must look beyond architecture, evaluating factors such as the availability of off-the-shelf software. Only then can they choose a winner from the many processor options.
Kenton Williston of BDTI contributed to this column.
Add new comment