BDTI recently completed an in-depth analysis of FPGAs' suitability for DSP applications. We found that, in some high-performance signal processing applications, FPGAs have several significant advantages over high-end DSP processors. Our recent benchmark results (shown in Figure 1), for example, have shown that high-end, DSP-oriented FPGAs have a huge throughput advantage over high-performance DSP processors for certain types of signal processing. And FPGAs, which are not constrained by a specific instruction set or hardwired processing units, are much more flexible than processors.
Figure 1. Results of the BDTI Communications Benchmark (OFDM)™
If market success were based solely on throughput or flexibility, FPGAs would appear to be on the verge of taking over the DSP market; in fact, according to a recent report from market research firm Forward Concepts, in 2005 Altera and Xilinx each had DSP FPGA revenues in excess of $200 million, selling more non-cell-phone DSP silicon than Freescale and Agere.
But of course, it's not that simple. Development effort, energy efficiency, cost-effectiveness, staff expertise, and market inertia (among other attributes) will all play a role in determining whether FPGAs become a dominant technology for DSP systems.
In this article, we'll share some of the key open questions that we've identified during the course of our analysis. These factors will affect FPGAs' success in DSP markets, and will be of significant interest to system designers who are considering using FPGAs in their signal processing systems.
Are FPGAs energy hogs, or not?
Energy efficiency is often a critical metric for signal processing applications. Battery-powered products are highly sensitive to energy consumption, and even line-powered products are often sensitive to power consumption, though it may be on a per-channel or per-unit-area basis. FPGAs have long been viewed as too power-hungry for most DSP applications, but we believe that this may be an obsolete perspective.
FPGAs use highly flexible architectures, and this flexibility is perhaps their greatest advantage. But flexibility comes with a hardware cost. More flexibility generally means more gates, more silicon area, more routing resources—and higher energy consumption. For this reason, FPGAs are generally less energy efficient than chips with dedicated hardware, such as ASICs.
But how do FPGAs compare to DSP processors? DSPs are highly tailored for efficient implementation of common DSP tasks, and thus many engineers assume that they are more energy-efficient than FPGAs. But DSP processors have their own inefficiencies. In a DSP, only a tiny fraction of the silicon is devoted to computation; most of the silicon area and most of the energy is devoted to moving instructions and data around. Hence, it would be a mistake to assume that FPGAs are inherently less energy efficient than DSPs.
In some high-performance signal processing applications, for example, FPGAs can take advantage of their highly parallel architectures and offer much higher throughput than DSPs. As a result, FPGAs' overall energy consumption may be significantly lower than that of DSP processors, in spite of the fact that their chip-level power consumption is often higher.
Unfortunately, there is a dearth of accurate, apples-to-apples energy consumption data for FPGAs and DSP processors, making it difficult to compare their energy efficiency As part of the analysis for our recent report comparing FPGAs to DSPs, "FPGAs for DSP, 2nd Edition," BDTI did its own "back of the envelope" comparisons of the energy efficiency of FPGAs and DSPs. Based on anecdotal data about FPGA power consumption, we estimated that high-end FPGAs implementing demanding DSP applications, such as that embodied in the BDTI Communications Benchmark (OFDM)™, consume on the order of 10 watts, while high-end DSPs consume roughly 2-3 watts. Our benchmark results have shown that high-end FPGAs can support roughly 10 to 100 times more channels on this benchmark than high-end DSPs, suggesting that their energy consumption per channel is significantly lower than that of DSPs. This contradicts the common view that FPGAs are energy hogs.
Obviously our comparison is based on very rough estimates. While we believe that FPGAs' energy efficiency is likely to be competitive—or even superior—to that of DSPs in many high-performance signal processing applications, this is still an open question. What's needed is a rigorous, well-controlled analysis of power consumption under comparable conditions.
Furthermore, the above analysis is only meaningful in the context of high-performance signal processing applications. Estimating the relative energy efficiency of FPGAs and processors for less-demanding signal processing applications would require a slightly different benchmarking approach and the evaluation of different chips (e.g., lower-power processors and smaller FPGAs); this is another area that's ripe for further investigation.