The rise of the digital signal processor
DSP architectures from Analog Devices and Texas Instruments continue to dominate military and aerospace applications, but a third device, the Motorola PowerPC, is giving chase because of its large installed base and easy programmability
DSP architectures from Analog Devices and Texas Instruments continue to dominate military and aerospace applications, but a third device, the Motorola PowerPC, is giving chase because of its large installed base and easy programmability
By John McHale n Keller
Digital signal processing systems and their applications are becoming more complex every day. As the needs for easy-to-use software development and high performance increase, military and aerospace systems designers find their choices most often lay among high-performance hardware from Analog Devices in Norwood, Mass., and software expertise and industry clout of Texas Instruments in Houston, Texas.
The flagship high-performance DSP from Analog Devices is the SHARC - short for Super Harvard Architecture Computer - a monstrously powerful floating-point machine that is quickly becoming the mainstay of military radar and sonar signal processing, and which is the heir apparent to the venerable Intel i860 DSP for defense applications.
Officials at Texas Instruments (TI), meanwhile, are known in military circles for their floating-point TMS320C40 family of DSPs. More recently TI designers are gaining attention for their fixed-point TMS320C62 DSP and are making waves with the announcement of their TMS320C67 fixed- and floating-point DSP, which will begin sampling later this year.
TI and Analog continue to lead in the military markets as leaders of Intel, Motorola, and Lucent Technologies slowly abandon their military business for the commercial world of telecommunications.
Still, the Intel i860 has a large installed base in military applications. Many users are satisfied with i860 performance, and are trying to find ways to keep the DSP in service for as long as possible, despite a lack of support from Intel. The Motorola PowerPC microprocessor, meanwhile, is showing impressive DSP performance in some military applications where the chip continues to carve out a niche of its own.
At present, TI holds 45 percent of the overall DSP market, and Analog holds 12 percent, says Will Strauss, president of Forward Concepts, a market research firm in Tempe, Ariz.
DSP design tradeoffs
Key issues confronting military DSP users include high performance, power budgeting, plastic and ceramic packaging choices, and code compatibility with legacy systems, explains Tom Smith, strategic marketing manager for DSP solutions at the TI Semiconductor military division in Midland, Texas.
Another big issue with DSPs is scaleability. "Applications need to be written so that they can be easily expanded," says James Wagget director of marketing at CSPI in Billerica, Mass. In addition, keeping up with new technology and avoiding obsolescence are important, Wagget says. "What a device needs is a red light to tell engineers when it is obsolete."
Wagget`s colleagues at CSPI offer DSP expansion with the MM-10 and MM-11 PCI mezzanine cards to provide DSP capability to single-board computers. The MM-10 offers four clustered SHARC processors and the MM-11 doubles that processing power with eight SHARCs.
"Obsolescence is becoming a serious issue due to the vanishing number of mil-spec components," says Joe Andrulis, vice president of marketing at Mizar Inc. in Carrolton, Texas. As a result, designers will be forced to move towards commercial-off-the-shelf (COTS) equipment and must be prepared either to change their environmental constraints or find other ways to ruggedize the components, he says.
Along with their questions about diminishing DSP lifecycles, military customers are asking if DSPs are industry standard, if they are they fast, and if they are scaleable, says Joe Sgro, president of Alacron in Nashua, N.H. Sgro`s engineers answer the scaleability question with their FT- Dominator-PCI computing subsystem based on the Analog Devices ADSP-2106x SHARC processor.
Designers will always push for greater chip performance and aim for more megaflops, more complex algorithms, and more channels with higher bandwidth, says Rodger Hosking, vice president of marketing at Pentek Inc. in Upper Saddle River, N.J.
Pentek engineers are getting increased performance from the Pentek Model 4290 Quad TMS320C6201 DSP, which uses the TI fixed-point TMS320C6201 DSP. With its advanced VelociTI architecture, capable of executing eight instructions in parallel, the device can produce 1.6 billion instructions per second at 200 MHz.
From a DSP, military customers are looking to squeeze the most processing power out of the least volume, says Barry
Isenstein, director of strategic marketing at Mercury Computer Systems in Chelmsford, Mass.
Yet while military DSP applications demand the highest performance, their market is not the main driver behind the DSP industry - the telecommunications industry is - points out Strauss of Forward Concepts.
That is a fair statement, says Len May of
Analog Devices. But Analog engineers design their technology for the military first, then create a lower-cost version for the commercial markets, he says. For example, the Analog Devices third-generation of the 21060 DSP receives funding from the U.S. Defense Advanced Research Projects Agency in Arlington, Va. The 21060 and 21062 still offer the most on-chip memory of any digital signal processor, May boasts.
The ADSP-21062 from Analog Devices builds on the ADSP-21000 DSP core to form a system-on-a-chip, adding a dual-ported on-chip static random access memory and integrated I/O peripherals supported by a dedicated I/O bus.
Fabricated in a high-speed, low-power CMOS process, the ADSP-2106x has a 25-nanosecond instruction cycle time and operates at 40 million instructions per second. With its on-chip instruction cache, the processor can execute every instruction in a single cycle.
Analog officials also recently released the ADSP-21065L the most powerful member of their SHARC family of 32-bit processors. It offers a performance 60 million instructions per second - or 180 million floating point operations per second.
For the majority of radar and sonar projects, SHARC is still the chosen DSP, says Mike Shadduck, technical director for Loughborough Sound Images in Leicestershire England. Its high performance and knack for doing fast Fourier transforms help make it popular, he says.
However, the choice of a DSP is usually not application driven, Shadduck continues. It is based on which DSP will run existing software.
In February 1997, Texas Instruments officials introduced their newest family of high-performance digital signal processors, the TMS320C6x. Based on TI`s VelociTI architecture, an advanced very long instruction word architecture for DSPs, the first device in this new generation was the TMS320C62x fixed-point DSP.
The `C6x generation now includes floating-point capability with the introduction of the 1-billion-operations-per-second TMS320C67x floating-point DSP core.
Leaders of the TI Semiconductor military division have released a military version of the 320C6201 with a performance of 1.2 billion fixed-point operations per second, and operating at approximately 150 MHz over the military temperature range of -55 to 125 degrees Celsius at about 5 to 6 Watts of power.
"We`re hearing from our customers that TI ran late with the delivery of the `C621," says a DSP software developer.
TI may have been late with some of the components by only a couple of months, but that happens with many breakthrough products on new silicon, says Pentek`s Hosking. "They didn`t have to wait long at all." Pentek began shipping their `C62-based 4290 DSP product last month, he says.
TI officials are also coming out with the 1 GFLOPS `C67x DSP core to complement their 1,600 MIPS `C62x core. With the C67x, TI will become the first DSP supplier to offer a compatible architecture for fixed- and floating-point processors. "Executable code run on the `C26x will be able to be run the same way on the `C67x," Hosking points out. Tentative scheduling has the C67x coming out in 1999.
TI experts intend to counter the SHARC`s advantages in raw performance by offering advanced software development tools. TI officials have increased their share of the DSP software market - and bolstered their software expertise and broad range of software offerings - with purchases of Tartan of Monroeville, Pa., Go DSP in Toronto, and most recently Spectron Microsystems of Goleta, Calif.
Spectron leaders will continue to support their SHARC customers, but eventually they will phase them over to TI products, says Liz Keith, strategic marketing manager at Spectron.
Supporting customers` software needs is an increasingly pressing issue among DSP suppliers. DSP hardware is becoming less expensive, while software is becoming more expensive, Hosking explains. As the performance keeps increasing and the DSPs become more complex, the demand for easy-to-use development software increases.
TI`s software development will improve their niche in the DSP market, says Mizar`s Andrulis. Many customers will follow TI just due to their market clout as the largest supplier of DSPs. Engineers at Mizar currently offer a the MZ 7772 Quad DSP engine, which provides the power of four 50 or 60 MHz TMS320C40 DSPs on a single-slot, 6U VMEbus card.
Mizar recently won a contract from Sanders, a Lockheed Martin Company in Nashua, N.H., to provide 79 DSP octal boards in support of the Joint SIGINT Avionics Family (JSAF) Low Band Subsystem - a new airborne intelligence subsystem designed to monitor a wide variety of communications. The system will initially be deployed on the U.S. Air Force RC-135, Army RC-7, and the Navy EP-3 reconnaissance aircraft.
At the end of last year Loughborough Sound Images merged with Mizar. The combined company will locate in Dallas.
Loughborough`s DBV42 and DBV44 modular DSP boards are part of a design at GEC Marconi in Cheshire, England, for a new mine-hunting simulation for the British Royal Navy. The DSPs use four TMS320C4x-based TIM-40 compatible module sites and are scaleable to more than 6 billion operations per second peak processing capability and have a scaleable topology.
Also offering fast DSP performance is Alex Computer Systems in Ithaca, N.Y., with the PAC80 Octal SHARCPAC mezzanine card, which operates at 960 million floating point operations per second, with throughput of 1.92 gigabytes per second for inter-processor communications. SHARCPAC is an industry-standard mezzanine card form factor designed to support several SHARC DSPs on each card.
Also using the SHARCPAC are engineers from Bittware Research Systems in Concord, N.H. Bittware officials offer SHARC technology with their SHARCPAC mezzanine modules, which provide modular, scaleable, and compact integration for multiprocessor DSP designs. SHARCPAC modules can add multiple processors and memory to platform-independent SHARPAC-ready carrier boards. The interface includes 48 data bits, 32 address bits, eight flag signals, eight interrupts, 16 link ports, and four serial ports.
Fixed point vs. floating point
Overall, floating-point DSPs are appropriate for processing dynamic, broad-ranging signals where designers cannot always predict signal parameters. Fixed-point DSPs, on the other hand, can offer faster performance than floating-point devices, but are appropriate for relatively narrow-range signals where systems designers always know the signal parameters. Exceeding the programmed signal range of a fixed-point DSP system - called "rollover" - has been known to cause catastrophic system failures.
The majority of military applications in radar, sonar, and mine detection today use SHARC floating point technology because of its ability to solve high-end computational problems with dynamic signal parameters, says Gerry Clancy, vice president of sales and marketing at Bittware.
In the floating-point world SHARC will continue to be king of the hill because of its I/O structure and high megaflops per second, says Mercury`s Isenstein.
Engineers at Mercury are offering the RACE Series MultiPort system, which comprises more than 1,000 SHARC and PowerPC processors in one system, and uses the RACEway Interlink switched fabric interconnect providing more than 12 billion floating point operations per second of processing power.
"For the military, fixed point is used mostly in secure communications, navigation, flight control, and airplane avionics, says TI`s Smith. Fixed point, meanwhile, is used in many image-processing applications for "pixel pushing," says CSPI`s Wagget.
While fixed point is cheaper on the component level, floating point saves money on the system level, says Richard Jaenicke, director of marketing at Sky Computers in Chelmsford, Mass.
"The selection of data format is a key design decision when choosing a DSP because different levels of hardware support are provided for floating point and fixed-point data types," says Graeme Harfman, product manager at Spectrum Signal Processing in Burnaby, British Columbia.
"A DSP`s data format determines its ability to handle signals-to-noise ratios," Harfman explains. "When considering the correct format to represent the data in an application, ease-of-use and time to market considerations are also equally important.
In general, real-world signals such as sonar, radar, and radio are better processed by floating-point DSPs," Harfman claims. "Real-world signals generally have a larger dynamic range than do constructed signals which makes them more difficult so scale." Fixed-point DSPs, on the other hand, are better suited to constructed signals such as in telecommunications because they generally have relatively low dynamic ranges and better signal-to-noise ratios, he says.
Experts at Spectrum offer the Fixed Point Monaco quad TMS320C6201 VME board, which they designed to feed data to the four `C6x processors at more than 1,700 megabytes per second each.
The PowerPC microprocessor has a growing presence in the DSP world. Not only does the device have respectable performance as a general-purpose processor and as a DSP, but it also can work easily with Windows NT, Unix, and other PC and workstation environments.
The PowerPC is a real wild card in the industry; it gives the traditionally hard-to-program DSP a link to the software-friendly world of Microsoft and Intel, Mizar`s Andrulis says.
Engineers at Ixthos are seeing an increased interest in the PowerPC due to its abilities in data analysis, says Ixthos President Jeff Milrod. Systems designers use the PowerPC mostly for low-end, non-real-time DSP solutions, he says.
The PowerPC`s ability to process general-purpose data and digital signals offers designers performance tradeoff choices that they rarely have had before. Dedicated DSPs provide performance, and PowerPCs provide compatibility with commercial technology and lower software and development cost, says Mark Russen, product marketing manager at Mizar.
By using dedicated DSPs together with PowerPCs in the same system, designers can use the DSP to handle high-end computational pre-processing, and the PowerPC to analyze the data, explains CSPI`s Wagget. CSPI brings to this market the MAP 2640, a quad PowerPC with four Myrinet interfaces and 64 to 256 megabytes on-board memory.
Using both devices brings together the commercial environment of the Power PC with the upside of DSP performance density, Mercury`s Isenstein says.
Ixthos`s Milrod, however, says he sees the market expanding overall for DSPs because of the need for high-end, complex solutions that only a dedicated DSP can successfully manage. Milrod`s engineers have developed the IXZ8 octal DSP product. This offering is the basis for the new conduction-cooled SVME/DMV-460 DSP board with eight SHARC DSPs in two clusters of four from Ixthos parent company DY 4 Systems of Kanata, Ontario. The Ixthos IXZ16 board offers 16 SHARCs in four clusters of four, with 1.9 GFLOPS on one 6U board.
Mercury engineers recently expanded their RACE Series to include systems based on Motorola`s new 300 MHz PowerPC 750 microprocessors. Called "Pontiac," these systems will be available in VME, PCI, and MultiPort heterogeneous multicomputer configurations. Mercury`s early tests indicate that "Pontiac" systems routinely provide twice the processing speed of previous PowerPC-based RACE systems for the embedded systems market.
Leaders of Sky Computers offer their Excalibur processor, which is based on the new PowerPC 740 microprocessor, the third-generation PowerPC. Each Excalibur includes four PowerPC 740 microprocessors and as many as 256 megabytes of dynamic random access memory. One Excalibur on a 6U VME motherboard delivers 2 billion floating point operations per second of performance.
Using the Excalibur are engineers at Lockheed Martin Tactical Defense Systems in Goodyear, Ariz., for their Foliage Penetrating Radar Program and by officials at the Telecommunications Research Institute in Warsaw, Poland, for their new mobile 3D radar systems.
The i860 from Intel Corp. in Folsom, Calif., used to be one of the most popular DSPs around and held a powerful niche in the military market, says Forward Concept`s Strauss. They may be missing out on some strong profit by abandoning it, he adds.
Although Sky Computers officials continue to offer the i860, they are trying to make it easy for their customers to migrate to the SHARC. "Intel will not be producing any new versions of the i860, but will continue to produce and support their current customers," Jaenicke says. "The i860 will stay around for a while simply because the military doesn`t let go of things that fast," Isenstein adds.
The architecture of the Monaco quad TMS320C6201 VME board from Spectrum Signal Processing in Burnaby, British Columbia, is designed to feed data to the four `C6x processors.
The Excalibur multiprocessor daughtercard from Sky Computers in Chelmsford, Mass., uses four PowerPC 740 processors and up to 256 MB DRAM.
The conduction cooled SVME/DMV-460 DSP board from DY 4 Systems in Kanata, Ontario, is based on the Ixthos IXZ8 octal DSP product and contains eight SHARC DSPs in two clusters of four.
New DSP core processes in fixed and floating point
Officials at Texas Instruments (TI) in Houston claim they will become the first suppliers to offer a digital signal processor (DSP) chip able to process floating-point, as well as fixed-point instructions.
The `C670, which is to complement their 1.6 billion-instruction-per-second TMS320C620 core, will begin sampling in the second half of this year, TI officials claim.
The first devices will provide 1 billion floating point operations per second of performance. Future generations of the device will be as fast as 3 GFLOPS by the end of the decade. The `C67x technology will cost less than $50, TI officials claim.
A planned military version of the `C670 is to run at 840 million floating point operations per second, TI officials claim.
The `C670 has eight processors on board - six of them floating point, and two of them fixed point. All `C620 instructions will run unmodified on the `C670 DSP, TI officials claim. "Now, executable code run on the `C62x will be able to be run the same way on the `C67x," says Rodger Hosking," vice president of marketing at Pentek Inc. of Upper Saddle River, N.J.
The `C670 can execute instructions in parallel, in serial, or in parallel/serial combinations. TI officials claim this scheme enables systems developers to reduce code size, power consumption, and the number of program fetches.
The `C670 core includes 8-, 16- and 32-byte addressability; 8 bits of overflow protection; saturation; bit-field extract, set and clear; bit counting; normalization, and two additional integer multiply functional units with 32-bit and 24-bit multiply support. - J.M.
Need for portable software drives DSP industry
Designers of digital signal processors (DSP`s) find the more complex DSPs become, the greater the need for industry-standard software to program and run the devices.
"The DSP industry is very fragmented," explains Mike Radhanauth, manager of the product marketing group at Spectrum Signal Processing in Burnaby, British Columbia. DSPs are used in so many different applications that it becomes difficult for a software standard to catch hold of the industry, he says.
Portability of code then becomes a deciding factor in choosing design software, Radhanauth says. Spectrum obtained software expertise last year by purchasing real-time-operating-system developer 3L Limited in Edinburgh, Scotland.
Engineers at 3L recently announced a version of their Diamond real-time operating system, which supports the TMS320C6201 DSP processor from Texas Instruments in Houston. Diamond is a multi-tasking, multi-threading real-time operating system for high-performance DSP systems.
Portability is not the only issue facing DSP systems developers when they choose software. They also need tools to help them program today`s ever-more-complex DSP hardware, says Liz Keith, strategic marketing manager at Spectron Microsystems in Goleta, Calif. Spectron produces the SPOX real time operating system, which uses a real-time, multitasking DSP kernel, host communications utilities, a function library, and source-code debugging.
Spectron engineers also offer the SPOXWorks, a software that converts C programs into executable DSP programs. It speeds algorithm development and deployment by giving designers access to real DSP hardware early in the design cycle. SPOXWorks performs all of its functions without requiring the user to write any lines of SPOX or DSP code.
Texas Instruments recently purchased Spectron as part of TI`s move to corner the market in DSP software. Other recent purchases by TI include GO DSP - makers of the debugging software Code Composer - and Tartan Inc., which specialized in software development tools for DSPs.
With the acquisition of Spectron, TI engineers now use the SPOX system with all of their DSP products, says Tom Smith, strategic marketing manager for DSP solutions at the TI military division in Midland, Texas.
Total DSP systems costs are mostly for software, and TI officials use their software expertise and market clout to drive this market, says Joe Andrulis, vice president of marketing at Mizar in Carrolton, Texas.
Engineers at The MathWorks in Natick, Mass., offer designers Simulink, a rapid-prototyping tool for modeling and simulating DSP systems and DSP blockset libraries. Simulink uses a block diagram editor to enable inexperienced programmers to design systems by selecting elements from the Simulink and DSP blockset libraries. MathWorks engineers also produce the Real-Time Workshop that converts Simulink and DSP Blockset into C programs.
Engineers most often choose new DSPs that are able to run their existing software, says Mike Shadduck, technical director for Loughborough Sound Images in Leicestershire England. Loughborough offers the IDE6000 graphic user interface-based DSP development environment for the TMS320C6201 from Texas Instruments.
Increasingly, engineers are willing to give up performance for portable code that they can re-use; the question is just how much they can afford, agrees Richard Jaenicke vice president of marketing at Sky Computers in Chelmsford, Mass.
Officials of White Mountain DSP in Nashua, N.H., say their customers want development tools that have plug-and-play compatibility and ease-of-use, says Jerry LaGrange, the company`s vice president of sales and marketing.
White Mountain engineers design the Mountain-510/WS Universal Workstation Emulator, which provides Sun Sbus-compatible C/assembly source level debugging for Texas Instruments DSP families.
Engineers at Ixthos in Leesburg, Va., and Alex Computer Systems in Ithaca, N.Y., work with the GEDAE software toolset from the Lockheed Martin Advanced Technology Laboratories in Camden, N.J. GEDAE supports the graphical development of DSP applications and the autocoding of the application for execution on workstations and/or embedded hardware.
Officials at Hewlett Packard Co.`s HP EESoft division in Palo Alto, Calif., offer the HP DSP designer software as part of their new HP Advanced Design System. The software includes a block-diagram algorithm-development environment, a DSP filter tool, a measurement instrumentation controller, and data post-processing capability.
The HP DSP designer`s built-in instrument controller helps design engineers avoid time-consuming and nonportable methods of hand-coding scripts by enabling them to interact with a user-friendly dialog box for data collection.
The Advanced Design System from HP EESoft, an electronic design automation software system, is the industry`s first integrated, end-to-end signal-path design solution for developers of communications products, claims Carter Smith, product specialist at HP EESoft. It places all the design elements into one common database, he says.
Scientists at CSPI in Billerica, Mass., used the message-passing middleware, MP-1 and MP-2, from MPI Software Technology Inc. in Starkville Miss., to work across different platforms. MPI`s technology enables engineers to run the same code on different computers.
Eonic Systems engineers in Herndon, Va., have a military and commercial product with their Virtuoso real-time operating system family, which uses Eonic`s SoftStealth technology to reduce code size. Eonic officials have just released Virtuso 4.0, a real-time operating system that runs on DSPs from Analog Devices and Texas Instruments.
Ada, the traditional programming language of the military, is losing its hold on the market and is being replaced by other languages.
"Ada is fading away; no one is starting out in Ada software anymore," Loughborough`s Shadduck says. It can be difficult to program and is not compatible with the many commercial systems that make up most of the DSP market, he says.
The military is still adjusting to the commercial-off-the-shelf issue, says Rick Hoffman, business development manager for Eonic Systems. Ada will continue to be used in mission critical and other legacy applications, he says.
While Ada was a great idea for its time, older military programs will slowly upgrade to using C, TI`s Smith says.
Officials at Analog Devices in Norwood, Mass., nevertheless, are pushing ahead with developing Ada tools for their SHARC DSP, officials say. The SHARC is often the DSP of choice for high-end military signal processing systems, and many military programs are still using the language, officials say. - J.M.
The IDE6000 software development tool from Loughborough Sound Images in Leicestershire, England, provides a GUI- based DSP development environment for the TMS320C6x from Texas Instruments.