High-Performance Fixed-Point DSP Processors
There are three mainstream competitors in the high-performance fixed-point arena: Analog Devices' ADSP-BF5xx (Blackfin), Freescale's MSC81xx and MSC71xx, and Texas Instruments TMS320C64x. All three are 16-bit fixed-point machines, and they are newer and more powerful than most of the chips in the low-cost group; for example, they all issue and execute multiple instructions using VLIW (very long instruction word) techniques, and they all support multiple MACs per cycle. (We should note that although there are only three processors in this group, there are many start-up companies that offer high-performance fixed-point DSP processors. These include multi-core devices and massively parallel architectures, and we'll cover them in a later article.)
Analog Devices ADSP-BF5xx
Analog Devices' ADSP-BF5xx (Blackfin) family combines features of low-power DSPs with features traditionally associated with general-purpose microprocessors, such as privilege modes and memory protection. It targets applications that require the functionality of both a DSP and a general-purpose processor (such as automotive applications), applications that require strong multimedia processing capabilities, and other computationally intensive applications.
Blackfin is based on the "Micro Signal Architecture," which was jointly developed by Intel and ADI. It has two data paths and implements a limited VLIW architecture—it can execute up to three instructions per clock cycle, and can perform two 16-bit multiplications per cycle. Pricing ranges from $5 to $32; because there are inexpensive chips in the family, Blackfin can also be classified as a low-cost DSP processor family, though its performance is significantly higher than that of the other chips in that group. The Blackfin family also includes multi-core chips.
The ADSP-BF5xx has an unusual and sophisticated power management scheme; ADSP-BF5xx family members are designed to operate over a range of clock speeds and operating voltages, and they include circuitry to ensure stable transitions between operating states. Blackfin isn't as fast as the other two chips in the high-performance category, but it is more power efficient. As a result, it's able to target portable, energy-sensitive applications in addition to wired high-performance applications.
Blackfin is notable because it is among the newer architectures that include specialized hardware and instructions to accelerate (two-dimensional) imaging and video applications. These capabilities represent a significant shift in DSP processor architectures, which were originally designed to process one-dimensional signals (such as in communications and audio).
Freescale MSC81xx and MSC71xx
Freescale's MSC81xx and MSC71xx families include chips that are based on three cores: the SC140, SC1400, and SC3400. The SC140 was introduced in 1999 and, along with Texas Instruments' TMS320C62x, it was one of the early mainstream VLIW-based processors, with StarCore and TI both promoting the VLIW approach as a way to make DSP processors more compiler-friendly. The SC140 is a quad-MAC architecture that can issue and execute six instructions per cycle.
The other two cores (the SC1400 and SC3400) are nearly identical to the SC140, but use different fabrication techniques and (in the SC3400) minor architectural changes to achieve higher performance. Of the three, the SC3400 is the fastest core, and it's currently only available in the MSC8144 chip. The MSC8144 contains four SC3400 cores that operate at 1 GHz. This chip targets multi-channel infrastructure applications such as wireless base stations, carrier-class voice-over-packet equipment, and media gateways, along with video surveillance applications.
The SC140 is used in a variety of single- and quad-core MSC81xx chips, the fastest of which execute at 500 MHz. MSC81xx chips target computationally demanding infrastructure applications, such as packet telephony media gateways, video multi-conferencing, and base stations. The SC1400 is used in single-core MSC71xx chips that operate at up to 300 MHz. These chips target lower-cost applications such as low-end IP telephony applications.
Chip pricing ranges from $13 to $84 for the MSC71xx/81xx chips, and from $180 to $220 for the quad-code MSC8144.
Texas Instruments TMS320C64x
Texas Instruments' TMS320C64x, introduced in 2000, is an extension to (and replacement for) Texas Instruments' earlier TMS320C62x. The TMS320C64x family targets high-performance applications such as wireless base stations, two- and three-dimensional imaging applications, video applications, and radar and sonar systems. TI recently announced some multi-core 'C64x-based DSPs that are intended for communications infrastructure applications. The fastest TMS320C64x family members execute at 1 GHz with a 1.2-volt core supply; single-core chip pricing is from $15 to $208.
The TMS320C64x has eight execution units, including two multipliers and four ALUs, and can execute up to eight instructions in a single clock cycle. The 'C64x also includes SIMD instructions that enable it to perform four 16-bit multiplications in parallel. (For comparison, the earlier 'C62x could compute only two parallel multiplications).
In 2005 TI introduced an enhanced version of the 'C64x architecture, the 'C64x+, which adds new instructions and can perform up to eight 16-bit multiplications per cycle.
The 'C64x+ operates at 1 GHz (like the 'C64) but by virtue of its additional instructions it is able to achieve noticeably higher DSP performance. Unlike the 'C55x+, the 'C64x+ is available in general-purpose DSP chips, and TI also uses the 'C64x/'C64x+ architecture in a variety of application-specific chips. These include the DaVinci video-oriented chips and the OMAP3430 application processor, which couples an ARM Cortex-A8 CPU with a 'C64x+ DSP.
|
