CMP - United Business Media TechOnline
All Articles Products Courses Papers VirtuaLabs Webinars Web



 
LoginRegister
      TechOnline > Learning >  Technical Paper
Under the Hood
November 18, 2002

Application Profiling: One Processor or Two?

Jeff Bier and Adam Lins
Berkeley Design Technology
TechOnline

For the developer of multimedia applications for handheld or embedded devices, the choice of type and number of processors has a significant impact on system design, cost, development time, and performance. Multimedia applications such as media players contain both signal-processing and control-oriented software components, and architectures that excel in one of these areas typically perform poorly in the other.

For example, architectures for DSP typically excel on signal processing workloads, such as video and audio encoders and decoders, and often do perform much less efficiently on control-oriented workloads such as operating systems, user interface control, and protocol stacks. The converse is generally true for general-purpose architectures.


Figure 1:  Clock speed required for real-time performance of an example multimedia application running on hypothetical general-purpose processor (GPP) and digital signal processor (DSP) cores.

Figure 1 illustrates the impact of these architectural characteristics for representative processors of each type, based on estimates of performance on a workload chosen from those typically found in a handheld multimedia application.

The workload includes: video encode and decode, MP3 decode, speech encode and decode, echo cancellation, operating system, player control and synchronization software, and user-interface control. Three scenarios are shown, corresponding to the choice of a standalone GPP (general-purpose processor) or DSP, or a multi-processor DSP plus GPP architecture.

Figure 1 shows that the DSP outperforms the GPP on signal-processing-intensive tasks, and the GPP outperforms the DSP on control-related tasks. In this example, choosing a multi-processor design and partitioning the application so that each processor is playing to its strengths reduces the processor clock speeds required to process the same workload in real time.

If the GPP alone were performing the tasks, it would need to be clocked at over four times the clock rate of the DSP+GPP combination to process the same workload in real time. A reduction in clock rate is valuable to the system designer: it may allow improved battery life, a reduction in weight or cost, or the possible addition of enhanced features.

What is not shown in Figure 1 are other important effects associated with multi-processor system design, for example, increased development effort arising from the increased complexity of the system design, and the overhead of inter-processor communication.


About the Authors
Jeff Bier is general manager, and Adam Lins is engineering manager, of Berkeley Design Technology, Inc. (BDTI). BDTI provides analysis of processors, tools, algorithms, and software used for DSP applications, and develops optimized DSP software. BDTI's Web site contains extensive free information of interest to DSP technology providers and users.

 
Rate this paper
WORSE | BETTER
1 2 3 4 5




   

TECH PAPER
1. Use Rowley CrossWorks and the MAXQ3120 Evaluation Kit to Create a Light Meter Application

TECH PAPER
2. System ACE Configuration Solution for Xilinx FPGAs

TECH PAPER
3. Interface Products Design Guide

TECH PAPER
4. Maintaining Data/Clock Synchronization with Spread-Spectrum EMI Reduction