Because DSPs are typically assigned MAC-intensive tasks, one might assume that DSP performance is directly related to MAC throughput. In fact, vendors often advertise the speed of their DSPs in terms of MMACS (millions of multiply-accumulates per second). However, BDTI benchmark studies reveal that MAC throughput is not a reliable indicator of real-world DSP performance.
One reason for this is that digital signal processing involves more than just MACs; e.g., Viterbi decoding—an increasingly important DSP task—doesn't use multiplication at all. And even MAC-intensive algorithms may not be able to make full use of a processor's MAC units—memory constraints often limit performance. The quad-MAC StarCore SC140, for example, can perform two 64-bit loads per cycle, but the loads must be aligned on 64-bit boundaries. Thus, if input data is unaligned, the SC140 cannot take full advantage of all its MAC units. Pipeline constraints may also limit performance: many instructions on the quad-MAC TI TMS320C64xx have multi-cycle latencies; this can significantly affect task execution times, particularly on short algorithms. Control overhead is also a key consideration—the StarCore SC110 and SC140 have better support for conditional execution than most DSPs, and thus these processors benefit from fewer inefficient test-and-branch sequences.
So while MMACS may matter, they are not everything—other factors play important roles in actual DSP performance. A recent BDTI study highlights this point: a 300 MHz single-MAC SC110 is actually about twice as fast as a 200 MHz dual-MAC TMS320C55xx. A more detailed analysis of this comparison is available in BDTI's new "Inside the StarCore SC110" report. Visit www.BDTI.com for more information.
Add new comment