Design Feature: August 3, 1995

In an application, the mantissa defines precision, and the exponent designates dynamic range. For example, in a missile-tracking system, each unit in the mantissa might represent an arc-second, and the exponent might indicate how many arc-degrees the entire tracking field included (how wide the field was overall).

Because fixed-point DSPs perform integer arithmetic, no exponent is included in the numeric format. Normally, fixed-point DSPs are either 16 or 24 data bits wide. A 24-bit fixed-point DSP offers the same precision as the 24-bit mantissa of a floating-point DSP. However, the fixed-point DSP does not offer the greater dynamic range of the floating-point DSP's exponent. Thus, with a fixed-point DSP, the missile-tracking system would be limited to a specific number of arc-seconds.

Consequently, the end use often determines the type of device a designer should choose. Factors that you should consider include power consumption, performance, programmability, packaging, and price.

The numeric format makes floating-point devices ideal for more complex operations in high-performance applications. However, the architectural differences give fixed-point devices a leading edge for high-volume, low-power applications.

Because of its 32-bit-instruction-word width, the floating-point DSP's instructions can be more powerful than the 16-bit instruction words of fixed-point DSPs. Although the MIPS ratings are higher for fixed-point DSPs, the 32-bit floating-point instructions support additional parallel operations.

Floating-point DSPs currently have a larger address width of 24 bits, compared with 14 or 16 bits for fixed-point DSPs. Fixed-point DSPs, which are used primarily in embedded applications, have a smaller address reach to reduce the number of external pins the application requires. Floating-point DSPs, which are optimized for system-level processing, are less sensitive to cost and require a larger address reach. Consequently, floating-point DSPs direct correspondingly larger areas of memory for greater data and program storage.

However, the floating-point DSP's greater word and address widths can add up to greater power consumption. Typically, the 16-bit fixed-point device's bus size allows it to fit into a smaller package and consume less power. You can operate the fixed-point device at faster speeds because of its relatively simple architecture and fewer speed paths.

Newer fixed-point DSPs provide application-specific instructions and on-chip power management for portable and mobile communications applications. Many designers are fine-tuning upcoming industry standards for 16-bit fixed-point implementations because of this implementation's prevalence in the mobile market. In addition, on-chip memories are fine-tuned to meet the needs of certain applications, such as digital cellular phones. The addition or elimination of certain peripherals, such as a host port, a UART, or a buffered serial port, is an option for the new fixed-point DSPs as well.

Traditionally, software development has been slower with fixed-point devices. However, breakthroughs in fixed-point design are improving the development-cycle time. For example, in a fixed-point implementation of a voice compression specification for the US digital cellular standard IS-54B, improved compiler support and a more orthogonal instruction set have greatly reduced development-cycle time for cellular phones.

Designers must weigh performance, power, and size with other important factors—device and system costs. Cost, of course, is a significant factor for many applications. In general, fixed-point DSPs tend to cost less. However, the application determines cost efficiency.

Table 1 highlights the differences between fixed- and floating-point DSPs. In general, fixed-point DSPs are lower cost devices. Designers, therefore, tend to use these devices in high-volume embedded applications. In contrast, designers use floating-point DSPs for system-level control in which performance, not cost, is the main concern.

Fixed-point DSPs are more peripheral-rich and include such functions as host ports, asynchronous and synchronous serial ports, buffered serial ports, and capture/compare. In addition, fixed-point DSPs with on-chip peripherals can rival floating-point processors in price, but fixed-point DSPs also help to save space and cost for other components. As a result, even the high-priced fixed-point devices tend to save on systemwide costs. With a greater peripheral and memory, fixed-point DSPs allow designers to match system requirements exactly.

Although the greater dynamic range and programming ease of floating-point operation come at a cost, some applications must have the higher floating-point format. The high-level language combined with the large address reach offered by the floating-point device make it ideal for applications such as workstations.

Table 2 lists the most important factors a designer must consider when deciding between a fixed- or floating-point DSP. The top of the table rates the DSPs from 0 (least favorable) to 5 (most favorable) for each of the design factors. The bottom of the table indicates typical applications in which these design factors are critical.

Because power consumption, cost, and size are benefits of fixed-point DSPs, these devices are obvious choices for cellular phones, modems, and hard-disk drives. However, for graphics and imaging, floating-point calculations and performance are indispensable, so these applications use floating-point DSPs.

Voice mail needs the low cost of a fixed-point DSP and the rapid development of a floating-point DSP. Voice-mail system developers may choose to use floating-point DSPs for low-volume, high-profit systems that handle 100 or more phone lines. The designers can redevelop the system using fixed-point DSPs for high-volume, low-cost systems that handle only a few phone lines.

Certain features of product selection and support may also affect your DSP choice. The additional development time of fixed-point operation may be cost-effective, depending on these features. Before choosing a DSP, you should investigate:

• The availability of software tools and development platforms;
• Whether a DSP architecture is available for the application. For instance, DSPs designed for audio/video decompression can be cost-effective for some consumer systems, although off-the-shelf DSPs may not be;
• The availability of off-the-shelf software modules for the application;
• Your ability to customize the DSP with added memory or logic for the application;
• The dynamic range that the application requires; and
• The packaging and parallelism requirements.

The choice of a fixed- or floating-point DSP for an application may not be a simple trade-off between cost and performance. The floating-point processor's dynamic range and ease of development, the fixed-point processor's simple architecture and on-chip peripherals, and your product selection and support affect your choice. A designer needs to explore all options to choose the right DSP.