May 8, 1997

EDN's 1997 DSP-Arichitecture Directory

Markus Levy, Technical Editor

Incredible things have happened in the DSP industry this year. But this statement would have applied last year, the year before that, and the year before that. This scenario implies that DSPs have become integral to many high-volume designs. As a result, DSP vendors are investing heavily in DSP architectures, more intelligent compilers, better debugging tools, and volumes of support software.

Vendors are focusing on several key aspects of new DSP architectures. The most obvious architectural improvements are in the increased "parallelism": the number of operations the DSP can perform in an instruction cycle. An extreme example of parallelism is Texas Instruments' C6x very-long-instruction-word (VLIW) DSP with eight parallel functional units. Although Analog Devices' Super Harvard Architecture (SHARC) can perform as many as seven operations per cycle, the company and other vendors are working feverishly to develop their own "VLIW-ized" DSPs. In contrast to superscalar architectures, VLIW simplifies a DSP's control logic by providing independent control for each processing unit. However, as TI demonstrates, the move toward VLIW architectures places more emphasis on compiler intelligence.

Speaking of compilers, new fixed-point DSP architectures continue to become more C-friendly. DSPs with software stacks, more flexible addressing modes, and more orthogonal registers and instructions adapt well to C. Several DSPs, such as Motorola's 568xx family, support a general-purpose µC-style in-struction set with bit-manipulation instructions that let you write control code without worrying about DSP complexities.

As DSPs take on more µC-like functions, the industry is also seeing µCs and µPs take on more DSP-like functions. However, DSP architectures are difficult to emulate. At the simplest level, µCs and µPs have been acquiring multiply-accumulate (MAC) units, but none performs single-cycle MACs. DSPs typically perform high-performance signal processing because they include parallel arithmetic-computation units, multiport register files, extended precision and dynamic range in the computation units, dual data-address generators, and instruction-looping hardware. Advanced RISC Machines (ARM) and Hitachi went for the gusto and added entire DSP appendages to their ARM and SH µPs, respectively. You can study the results of performing digital signal processing on µPs by checking out Berkeley Design Technology's (Fremont, CA) report, "DSP on general-purpose processors."

Another way for you to add DSP functionality to a µP or any system is by using DSP cores. These cores are taking on a life of their own. The DSP Group has now acquired 18 licensees for its cores. Analog Devices has also delivered some of its cores to several licensees. Or, if you don't want a full-featured DSP, you can buy any number of function-specific cores from companies such as Technical Data Freeway (San Diego) and Mentor Graphics (Wilsonville, OR). These functions, which you can plug into ASICs and FPGAs, range from FIR filters to image processors.

Cores provide one means for you to add signal processing to a system. But signal processing has become much more than just a silicon product. Dozens of companies, such as HotHaus (Richmond, BC, Canada), DSP Software Engineering (Bedford, MA), and Analogical Systems (Palo Alto, CA), provide finely tuned software algorithms for a variety of applications. Furthermore, companies that have traditionally supplied DSP boards, such as Spectrum Signal Processing (Burnaby, BC, Canada) and Pentek Inc (Norwood, NJ), now also sell application suites of software that support their boards in key application areas.

Without further delay, it's time to see how the DSPs have multiplied. Shift into gear as this DSP directory guides you through the latest accumulation of MAC units.

| EDN Access | Feedback | Table of Contents |