Texas Instruments Announces $5.75 Floating-Point DSP

In September, Texas Instruments, Inc. (TI) announced the TMS320C6720 floating-point DSP. Operating at 200 MHz and priced at a modest $5.75 (unless otherwise stated, all prices mentioned in this article are for 10,000 unit quantities), the chip is the latest and least expensive member of TI’s ‘C67x family of 32-bit floating-point DSPs. It is intended for cost-sensitive applications such as musical instruments, medical imaging and biometrics.

The ‘C6720 is pin-compatible with the ‘C6722 and ‘C6726 DSPs, and like those chips features a 32 KB instruction cache and 384 KB ROM. Unlike the ‘C6722 and the ‘C6726, however, the ‘C6720 includes only 64 KB of on-chip non-cache RAM (a reduction from 128 KB and 256 KB for the ‘C6722 and ‘C6726, respectively).  The smaller on-chip RAM will detract from performance in memory-intensive applications. Memory differences aside, the ‘C6720 should perform identically to the ‘C6722 (200 MHz, priced at $9.95) and be about 25% slower than the ‘C6726 (250 MHz, priced at $14.10). For applications that are relatively insensitive to on-chip memory capacity, the ‘C6720 thus offers a much better price/performance ratio than the older ‘C6722 and ‘C6726.

Floating-point DSPs are experiencing a renaissance of sorts. In the late 1980s and early 1990s there was considerable investment in the development of floating-point DSPs by vendors like Analog Devices, Motorola (now Freescale), AT&T (now Agere), and TI, targeting high-end applications such as radar and medical imaging.  But in the mid-1990s, DSP chip vendors’ focus largely shifted towards fixed-point processors aimed at higher-volume markets. High-performance general-purpose processors became faster in the meantime, and added DSP capabilities; they replaced floating-point DSPs in many high-end applications. Lately, however, there has been increasing interest in low-cost floating-point DSPs for a variety of applications, particularly high-fidelity audio, with both TI and Analog Devices, Inc. (ADI) introducing new products on a regular basis.

In many signal processing application areas, algorithms are first developed with floating-point data types, typically using C or MATLAB models.  If these algorithms are to be implemented on a fixed-point chip, careful analysis and experimentation is usually required to find the selection of fixed-point data types and the combination of scaling, rounding, and saturation operations that achieve the necessary numeric fidelity with a reasonable computation load.  Using a floating-point DSP allows developers to bypass this step, reducing development effort and shortening development schedules. Despite these advantages, historically their high cost has limited the deployment of floating-point DSPs in cost-sensitive applications.

In November of 2005, ADI announced the 266 MHz ADSP-21375 floating-point chip.  Priced at $7.95 the chip targets many of the same applications targeted by the ‘C6720. ADI has also subsequently decreased the price of its 150 MHz ADSP-21261 floating-point DSP to $5.25. BDTI benchmark scores are not yet available for the ‘C6720, but BDTI estimates that the 200 MHz ‘C6720 will be faster than the 150 MHz ‘21261 but slower than the 266 MHz ‘21375 on typical signal-processing tasks.

It is also interesting to compare the ‘C6720 to its fixed-point counterparts. Floating-point DSPs, owing to their more-complex arithmetic hardware and larger data word sizes, tend to be more power- and memory-hungry, and have traditionally been more expensive than their fixed-point counterparts. But the ‘C6720 is less costly than TI’s 200 MHz ‘C5502 ($6.32) which offers similar performance and on-chip memory, ignoring the floating-point vs. fixed-point difference.

The TMS320C6720 is sampling now. TI also offers ‘C672x developers a Professional Audio Development Kit (PADK), consisting of the ‘C6720 integrated with ADCs, DACs, and a host of DSP software development tools, for $1,995.