Determining what to optimize depends on the goals of the application developer. For example, optimizing for performance means that a developer can use a slow or less expensive DSP to do the same amount of work. In some embedded systems, cost savings like this can have a significant impact on the success of the product.

The tricky part to optimizing DSP applications is to understand the trade-offs between the various performance parameters such as throughput (execution speed), memory usage, I/O bandwidth and power dissipation.

For instance, optimizing an application for speed often means a corresponding decrease in power consumption, but an increase in memory usage. Optimizing for memory may also result in a decrease in power consumption due to fewer memory accesses, but an offsetting decrease in code performance. The various trade-offs and system goals must be understood and considered before attempting any form of application optimization.

The fundamental rule in computer design as well as programming real-time DSP-based systems is "make the common case fast and favor the frequent case." This is really just Amdahl's Law that says the performance improvement to be gained using some faster mode of execution is limited by how often a developer uses the faster mode of execution. In other words, developers shouldn't spend time trying to optimize a piece of code that will hardly ever run. They won't get much out of it, no matter how innovative they are. Instead, if developers can eliminate just one cycle from a loop that executes thousands of times, they will see a bigger impact on the bottom line.

DSP architectures are designed to make the common case fast. Many DSP applications are composed from a standard set of DSP building blocks such as filters, Fourier transforms and convolutions. These algorithms all share a common characteristic; they perform multiplies and adds over and over again. This is generally referred to as the sum of products (SOP). DSP chip designers have developed hardware architectures that allow the efficient execution of algorithms with SOPs. This is done using specialized instructions such as single-cycle multiply and accumulate (MAC) architectures that allow multiple memory accesses in a single cycle and special hardware that handles loop counting with very little overhead.

DSP algorithms can be made to run faster using techniques of algorithmic transformation. For example, a common algorithm used in DSP applications is the Fourier transform. The Fourier transform is a mathematical method of breaking a signal in the time domain into all of its individual frequency components. The process of examining a time signal broken down into its individual frequency components is also called spectral analysis or harmonic analysis.

There are different ways to characterize Fourier transforms: