The question of which technology is most appropriate is age-old. When I made a presentation on this topic at the International Conference on Signal Processing Applications & Technology in 2000 the session was standing room only. My paper was inspired by analyzing how I made the decision as to which technology was appropriate (you can find my paper here: http://www.techonline.com/electronics_directory/techpaper/193103740;jsessionid=T3S5J3SJGT4WKQSNDLRSKHSCJUNN2JVN). It dawned on me that my decision depended solely on the attributes of the end product and not on attributes of the technology used.
I identified four end-product attributes that, when used in combination, discriminated between solutions. The attributes are: Speed, Math, Decisions, and Complexity. Pick the most important of these attributes of your product and I can tell you the most appropriate technology.
Micros shine in Decisions and Complexity. They have huge memory reach to store incredible repertoires of functionality that is switched with the speed of a branch instruction. FPGA's shine at speed and math. Why not? Entire algorithms can be parallel-wired and executed in a single cycle with proper pipelining. DSPs are decent across-the-board performers and not the best in any one dimension. The DSCs marginally improve DSP performance in decisions and complexity but are not usually the fastest DSPs.
About the time I made that presentation, the hybrid microcontroller-in-FPGAs were becoming available. Altera had announced its Exalibur and NIOS-based devices. Atmel already had introducted its FPSLIC. I reasoned that these SoC solutions would be winners. Little did I imagine how powerful and popular the soft-cores would become!
As time passed, some engineers argued that a DSP is better than I said. They said a DSP was built for math and can't be anything less than the highest performer. I listened and considered their point and then realized a salient point I wish I made.
For certain, an FPGA can outperform a DSP in any single math application. However, two factors mitigate choosing an FPGA over a DSP for math. One is whether the math is needed 100% of the time. The real estate of an FPGA is expensive if not always used. There is often a way to pipeline the processing to make the DSP adequate. Second, if the number of math functions is large. It is time consuming to repurpose an FPGA. A DSP, like a micro, can be repurposed for different math functionality with the simplicity of a branch.
The question of which technology is most appropriate is age-old. When I made a presentation on this topic at the International Conference on Signal Processing Applications & Technology in 2000 the session was standing room only. My paper was inspired by analyzing how I made the decision as to which technology was appropriate (you can find my paper here: http://www.techonline.com/electronics_directory/techpaper/193103740;jsessionid=T3S5J3SJGT4WKQSNDLRSKHSCJUNN2JVN). It dawned on me that my decision depended solely on the attributes of the end product and not on attributes of the technology used.
I identified four end-product attributes that, when used in combination, discriminated between solutions. The attributes are: Speed, Math, Decisions, and Complexity. Pick the most important of these attributes of your product and I can tell you the most appropriate technology.
Micros shine in Decisions and Complexity. They have huge memory reach to store incredible repertoires of functionality that is switched with the speed of a branch instruction. FPGA's shine at speed and math. Why not? Entire algorithms can be parallel-wired and executed in a single cycle with proper pipelining. DSPs are decent across-the-board performers and not the best in any one dimension. The DSCs marginally improve DSP performance in decisions and complexity but are not usually the fastest DSPs.
About the time I made that presentation, the hybrid microcontroller-in-FPGAs were becoming available. Altera had announced its Exalibur and NIOS-based devices. Atmel already had introducted its FPSLIC. I reasoned that these SoC solutions would be winners. Little did I imagine how powerful and popular the soft-cores would become!
As time passed, some engineers argued that a DSP is better than I said. They said a DSP was built for math and can't be anything less than the highest performer. I listened and considered their point and then realized a salient point I wish I made.
For certain, an FPGA can outperform a DSP in any single math application. However, two factors mitigate choosing an FPGA over a DSP for math. One is whether the math is needed 100% of the time. The real estate of an FPGA is expensive if not always used. There is often a way to pipeline the processing to make the DSP adequate. Second, if the number of math functions is large. It is time consuming to repurpose an FPGA. A DSP, like a micro, can be repurposed for different math functionality with the simplicity of a branch.
