AltiVec continues its charge up the DSP hill
The AltiVec general-purpose processor from Motorola is eating into the digital signal processor market share of Analog Devices and Texas Instruments, yet the two companies' chips have long-term roadmaps and continue to expand their share of the embedded market
By John McHale
Designers of digital signal processor (DSP) boards for radar, sonar, and signal-intelligence applications continue to embrace Motorola's PowerPC 7400 microprocessor, also known as the AltiVec. Customers are demanding the reduced instruction set computer's (RISC) 14.4 billion floating point operations per second (gigaflops) of performance.
Leaders of Mercury Computer Systems and Sky Computers, both of Chelmsford, Mass., have practically abandoned producing dedicated DSP boards for new design wins and deal only with the AltiVec.
The traditional DSP chip providers, Analog Devices in Norwood, Mass., and Texas Instruments (TI) in Houston, are not riding into the sunset anytime soon. Both companies offer DSPs with technology roadmaps targeting even higher levels of performance than the AltiVec from Motorola Semiconductor Products Sector in Austin, Texas.
"AltiVec is encroaching on the floating point DSP market due to sheer performance," explains Doug Patterson, director of marketing at Ixthos in Leesburg, Va. The RISC chip not only can pull data from outside memory but it can also perform Fast Fourier Transfers (FFTs) in about 20 to 25 microseconds, he adds.
The 64-bit AltiVec processor does an exceptional job with adaptive beamforming applications such as sonar and radar, says Bernard Pelon, director of market research at CSPI in Billerica, Mass.
"We can get improved performance out of an AltiVec and it is easier to program," says Steve Paavola, marketing manager at Sky Computers. "Our customers want the AltiVec," therefore Sky's [Analog Devices] 21060 products have been discontinued. The company, however, will continue to support their legacy customers as long as parts are available, Paavola adds.
PowerPC popularity stems from its development environment, which is friendlier than those for DSP, says Ron Marcus, marketing director at Synergy Microsystems in San Diego. It comes down to deciding which environment you want to program in, he adds.
The Merlin Altivec DSP board from Sky Computers is part of a British radar signal processing system from BAE Systems.
The Sky Computers AltiVec-based Merlin products are being used in a British radar signal processing system from BAE Systems in England. Each Merlin includes four 333 MHz MPC7400 PowerPC microprocessors that designers can configure with 1 gigabyte of on-board memory. The Merlin 6U VME board delivers 10.6 gigaflops of performance with total performance in one VME chassis of 672 gigaflops.
Several vendors of fast and powerful boards, such as Mercury and Sky, use PowerPC for its software advantages, says Roger Hosking, vice president of marketing at Pentek Inc. in Upper Saddle River, N.J. In large multi-processing systems, programming DSP chips and working link ports can be very complicated, Paavola says.
For this reason, company leaders must consider the cost of hiring and keeping competent software engineers, CSPI's Pelon says. Many engineers today prefer working with RISC processors because they are easy to program, Pelon adds.
The PowerPC architecture also works well with operating systems such as VxWorks from Wind River Systems in Alameda, Calif., Pelon says.
CSPI's 2840 multicomputer uses AltiVec Technology combined with Myrinet gigabit-per-second network technology for applications such as radar, sonar, and simulation. The device is designed with an 8-port crossbar switch to interconnect the four processing nodes on the board and allows for additional off-board communication.
Each processing node consists of two levels: a firstlevel for managing network communications, and a second level for compute processing.
Mercury engineers have found that the RISC chips have more forward and backward compatibility than floating-point DSP devices, says Richard Jaenicke, director of product marketing at Mercury. This is so because DSPs have longer life spans than do microprocessors. One frustrating aspect, however, is that original SHARC products are not forward compatible with the Analog Devices new TigerSHARC device, Jaenicke explains.
This is one of the reasons that Mercury designers have decided not to use the TigerSHARC, he adds. Also, "the device's performance per watt isn't compelling enough for us to choose it," he says.
In radar, sonar, and signal-intelligence applications, RISC will continue to replace DSPs because RISC is more powerful, easier to use, and has backward and forward compatibility, which helps combat obsolescence, Jaenicke claims.
The PowerPC is starting to move ahead of the SHARC in terms of performance, Jaenicke says. Mercury engineers will use RISC-based chips in most of their new designs, he adds. The performance boost starts to make a difference "when you get a two to three year difference in ease of use," Jaenicke explains.
Lovers of TI DSP chips will need to see RISC have five times the performance before they even consider switching, Jaenicke says. Their loyalty is strong, he adds.
The TigerSHARC digital signal processor chip from Analog Devices has a 64-bit external data bus and an internal bandwidth of 7.2 gigabytes per second.
Mercury engineers offer their RACE++ series VME systems that combine the Motorola AltiVec microprocessors running at 400 MHz with SDRAM memory operating at 133 MHz. This improvement in memory frequency enables a speed increase of as much as 60 percent for today's memory-bound signal processing algorithms.
"Many digital signal and image processing applications are limited by memory bandwidth," Jaenicke says. "By boosting the memory speed to 133 MHz while simultaneously boosting the processor speed to 400 MHz, we are delivering the optimum mix of processor, memory, and key technology components in a balanced system configuration."
DSP chip designers admit the AltiVec has some nice features, but they say the device does not specifically focus on signal processing the way a DSP chip does.
Although multiprocessing companies like Mercury and Sky have switched from SHARC products to PowerPC, Analog Devices and TI engineers still offer powerful alternatives to RISC technology.
Yes, Mercury and Sky like the AltiVec's links to operating systems, but TigerSHARC is ideal for multiprocessing because it performs its functions on one chip, whereas the AltiVec takes three chips to do the job, claims Colin Duggan, product marketing manager for Analog Devices.
TigerSHARC does not require extra circuitry like the memory controller and C2 cache that the AltiVec needs, Duggan says. For a radar or sonar application where power, size, and cost are optimum "you have to go with TigerSHARC," Duggan says.
In many key areas TigerSHARC comes out on top when compared to the AltiVec chip, Duggan claims. Analog Devices officials list one 1 Ku SHARC as $210, operating at 5 watts, and covering 1.1 square inch in area, while they estimate the AltiVec PowerPC 7400 to be $375, running at 17 watts (this number includes the AltiVec chip, C2 Cache, main memory, and the memory controller), and 13.7 square inches in area.
Duggan and his colleagues also estimate that when performing a 1,024 point complex FFT, the TigerSHARC's megaflops per watt is 112, while the AltiVec PowerPC 7400 is only 25, Duggan says.
The performance of the TigerSHARC is superior and its clock rate can skyrocket, says Graeme Harfman, business development manager at Spectrum Signal Processing in Burnaby, British Columbia. Performance per watt on the AltiVec may be better in an individual board but in terms of an overall system, the TigerSHARC comes out on top, Harfman claims.
AltiVec processors are also not as good for digital radio applications as a traditional DSP chip, which is more targeted at handling real-time processing, claims Mike Radhanauth, director of business development and strategic marketing at Spectrum Signal Processing.
Duggan admits the TigerSHARC is a different architecture from previous versions such as the Analog Devices Hammerhead. The new architecture enables Analog Devices engineers to achieve increased performance with 16 bit math functions, he says.
Other TigerSHARC features include a 64-bit external data bus with built-in SDRAM or burst SRAM controllers, an internal bandwidth of 7.2 gigabytes per second, 14 direct-memory-access channels for link port and external memory to internal memory transfers, four bi-directional link ports with 600 megabytes per second throughput, 6 megabytes of on chip memory, and the device supports as many as eight processors in a cluster multiprocessing system.
The TigerSHARC roadmap shows an increase of three to five times in performance over the next few years, Duggan says.
TigerSHARC also has a different instruction set than Hammerhead, but both use the Analog Devices VisualDSP development tool. Also, is not difficult to migrate from Hammerhead to TigerSHARC, Duggan explains.
Analog Devices engineers have gained ground in the software development area with their VisualDSP tool, Radhanauth says. The tool is now in version 4.1.
Analog Devices engineers also recently added the C++ language to their VisualDSP development tools. "With increasing time-to-market demands, many new DSP applications are using a higher percentage of C programming language versus DSP assembler," says Geoff Millard, compiler manager for DSP tools at Analog Devices. "Our implementation of C++ takes this step further by providing easier access to specialized features in DSP architectures."
For a fast real-time environment DSP is still the best choice, agrees Tom Smith, strategic marketing manager for military products at TI.
TI engineers recently released their TMS320C6203 DSP, which is the chipmaker's latest TMS320C6x family product offering. The chip uses TI's new 0.15-micron process, which delivers 896 kilobytes of on-chip SRAM, 512 kilobytes for data and 384 kilobytes for program, which eliminates the need for external memory access in many applications. The DMA controller transfers data to peripherals over dual 32-bit external data busses at rates as fast as 600 megabytes per second. The performance upgrade path is the same package as the TMS320C6201, Smith adds.
PowerPCs are general-purpose processors and perform all general functions including some traditionally done by DSP chips, Smith says. However if you have a special need for hard real-time signal processing you go to a specialist, which is a DSP chip, Smith explains.
In some cases DSP and RISC chips can coexist on the same board, Smith says. That is a possibility, Analog Devices's Duggan agrees. DSP would perform the hard real-time signal processing and PowerPC would handle the other functions, he says.
Motorola engineers designed the AltiVec as a general-purpose processor, Pentek's Hosking says, and the fact that it can perform DSP functions is icing on the cake. The SHARC and TI's TMS320 products still are the choice for hard real-time signal processing functions, Hosking says.
TI officials are targeting cellular phones with their fixed-point chips and Analog Devices is aiming at high-volume heavily embedded applications with their $10 DSP chips, Hosking explains. DSP chips are not in any jeopardy, Hosking adds.
The life cycle of DSP chips is longer than that of RISC processors which is why DSPs are attractive to the military, says Jerry Clancy, vice president of sales and marketing at Mango DSP in Norwalk, Conn. DSPs generate less power and take up less space which makes them ideal for avionics applications, he adds.
TI is still winning contacts for its TMS320C31 products, specifically on the U.S. Air Force F-16 and F-22 jet fighters, Smith says.
Many of the Analog Devices military customers are also concerned with obsolescence issues, Duggan says. DSP chips do not go obsolete as fast as RISC chips, he claims. "We still support our first SHARC product, the 21020," Duggan adds.
"DSP chips are still lower power and if you don't need external memory it is a less expensive solution," Sky Computers' Paavola notes.
TI officials also recently announced their TMS320C64x, which will have clock speeds of 1.1 gigahertz and performance near 9 bilion instructions per second. This delivers 10 times the performance of TI's TMS320C62x. The C64x DSP core is software-compatible with TI's previous DSP generations for faster time to market and reduced development time, TI officials say.
Ixthos officials are marketing their VME CHAMP-AVa DSP board, which is based on Motorola's AltiVec chip and can function as a rugged, conduction-cooled digital radio.
The CHAMP-AV (common heterogeneous architecture for multiprocessing) "is ideal for tactical and strategic radio programs for joint services," Ixthos's Patterson says. The board's high-throughput concurrent multiprocessing architecture can run at 14.4 billion floating point operations per second (GFLOPS) in one 6U VME slot. The device provides support for two PMC expansion modules, which enables the board to function as a digital radio, Patterson adds.
Spectrum Signal Processing engineers have designed the PEM-4WDC for digital radio applications. It is a wideband digital down converter module based on Spectrum's PEM specification. The PEM-4WDC accepts two digitized IF streams from Spectrum' s PMC-2MAI ADC module at an input sample rate as fast as 65 MHz. The IF streams are transferred from the PMC-2MAI to the PEM-4WDC on a ribbon cable using low voltage differential signaling for low power consumption and high noise immunity.
The PEM-4WDC has a synchronization bus that is used for synchronizing multiple down converters. A front panel connector provides access to the synchronization bus so that multiple PEM-4WDC modules can also be synchronized.
The PPC4 and PPC4a single-board computers from Radstone Technology in Towcester, England, are based on the AltiVec MPC7400 PowerPC processor. The boards run at 400MHz with an 83.3 MHz memory bus to yield sustainable memory transfer rates of 190 megabytes per second and 296 megabytes per second respectively. Other features include dedicated Level 2 Cache support, a range of DRAM options, flash capacity, 100BaseT Ethernet support, and two PMC slots.
Pentek engineers have developed the Model 4292 VME single-board computer. The board is designed for applications requiring extremely high-speed signal processing and uses four TI TMS320C6203 DSPs. Operating at 300 MHz, they deliver a combined peak processing performance of 9,600 MIPS.
Synergy engineers use the Motorola AltiVec chip with their VGM5 single-board computer. The VGM5 boosts data throughput with Synergy's VME Speedway direct CPU-to-VME data path. The board's P-zero-PCI bus enables connection of multiple Synergy VGR5s with a high-speed PCI bus using an existing VME64x backplane. The board also offers the choice of VxWorks or single-processor Linux support.
The TPE3 designed by experts at Transtech DSP in Ithaca, N.Y., is a high performance PCI card that uses the MPC7400 PowerPC processor. The device is aimed at telecommunication, simulation, and control applications. The card also has as many as 2megabytes of L2 cache and 512 megabytes of SDRAM.
GEDAE development tool flies with Merlin on P-3 Orion
CAMDEN, N.J. - The GEDAE DSP development tool from Lockheed Martin Advanced Technology Laboratories has been ported to the Merlin AltiVec PowerPC from Sky Computers in Chelmsford, Mass., and is flying with Merlin on the U.S. Navy P-3 Orion aircraft.
"Our customers like GEDAE because it provides graphical orientation for developing applications" and works rapidly between processors, says Steve Paavola, marketing manager at Sky Computers.
Other board vendors find their customers are asking for GEDAE as well.
"GEDAE provides 2000 SERIES customers another valuable tool simplifying application development for scalable, multiprocessor, real-time systems," says James Waggett, vice president of application development at CSPI in Billerica, Mass.
The GEDAE DSP development tool from Lockheed Martin Advanced Technology Laboratories is flying on the U.S. Navy P-3 Orion aircraft.
One customer who just started using GEDAE reduced his development time by a factor of three and expects to reduce it by a factor of five to ten over the life of the product, says Paul Houlihan, GEDAE sales manager for Lockheed Martin Advanced Technology Laboratories. Lockheed officials want to make GEDAE the development tool of choice for commercial-off-the-shelf vendors, Houlihan adds.
Lockheed engineers also recently released version 3.1 of GEDAE with a number of improvements over the 3.0 version, Houlihan says.
Many of the new features were the result of user suggestions, Lockheed officials say. In addition a large number of bugs have been fixed and the internal schedulers are significantly faster, Lockheed officials claim.
One of the improvements directly affected CSPI customers, Lockheed officials say. The direct schedule access communication capability for CSPI hardware has been extended to support multiple buffers on both the send and receive side of the communication, they explain.
The new features for GEDAE include benchmarks defined by Mitre Corp. The benchmarks include a corner turn, 2-D Fast Fourier Transforms, synthetic aperture radar, and space-time adaptive processing. In addition to benchmarking the performance of GEDAE, the benchmarks test the operation of the development tool on vendor hardware and are good examples of implementing common types of processing.
Lockheed engineers have implemented into GEDAE 3.1 the ability to hold output memory in a function box, which reduces the amount of memory used and the number of data copies required.
Version 3.1 also includes equation-based parameter settings to enable developers to use an equation to set the parameter values for all family members with a single operation.
Other additions include improved graphic execution, a multihop communication capability that enables GEDAE to distribute a graph over a complex architecture, and extending the GEDAE command program interface to enable setting the partition to processor mapping from a command program.
For more information on GEDAE contact Lockheed Martin Advanced Technology Laboratories by phone at 609-338-3452, by fax at 609-338-4166, by mail at 1 Federal Street A&E Building 3W Camden, N.J. 08102, or on the World Wide Web at http://www.gedae.com. - J.M.
Catalina delivers 133 MHz with Pathfinder-2
COLORADO SPRINGS, Colo. - Engineers at Catalina Research recently released their Pathfinder-2 high-performance digital signal processor aimed at radar, sonar, and signal intelligence applications.
Catalina focuses on low-latency, general-purpose frequency-domain applications, says Michael Bonato, director of marketing at Catalina. The Pathfinder-2 performs vector processing and excels at Fast Fourier Transforms (FFTs) and polyphase filters, he adds.
The new device does more in less time than previous versions of Pathfinder, Bonato says. Pathfinder-2 has real-time sustainable performance that is necessary for doing wideband spectral analysis in radar applications, he adds.
The Catalina device performs a 1K FFT in "15.4 microseconds at a sustained rate over and over and over," Bonato claims. It runs 5.3 billion floating point operations per second, while generating 3 watts of power, he says. The device has a 3.3 volt I/O and a 2.5 volt core.
Bonato says that the Pathfinder-2 does not compete with the AltiVec PowerPC7400 because the Pathfinder-2 is up front, while the AltiVec is in the back end.
Its high precision and handling of internal scaling enables Pathfinder-2 to process large vector sizes (as many as 1 million complex samples), Catalina officials claim.
Pathfinder-2 provides a multi-port data flow structure designed to support concurrent I/O and processing, making the chip a good match for applications requiring very fast data throughput rates, Catalina officials say. Multiple Pathfinder-2 DSPs can also be cascaded for increased performance.
The Pathfinder-2 integrated circuit is built up from multiple complex multiplication stages and four radix-four cores. It can accept and produce data as either 32 bit integer or IEEE floating point format. Pathfinder-2's internal arithmetic elements use floating point adders and multipliers for increased range and precision. The internal floating point element's word width has been optimized for speed, area, and precision, Catalina officials say.
Pathfinder-2 also uses an internal sine/cosine (twiddle) generator. This generator enables Pathfinder-2 to compute up to 1 million point complex or 2 million point real FFTs with no external twiddle coefficients required.
The device also provides five bi-directional I/O ports. The chip allows full crossbar multiplexing on all five ports, enabling flexible system designs and algorithm implementation. - J.M.
Greek air force airborne radar to use Mercury RACE systems
CHELMSFORD, Mass. - Engineers at Ericsson Microwave Systems AB of Molndal, Sweden, are using RACE signal processing systems from Mercury Computer Systems for their ERIEYE Airborne Early Warning and Command radar system to be supplied to the Greek air force for installation on a fleet of four EMB-145 jet aircraft.
The ERIEYE system detects, tracks and manages airborne, surface and low-flying targets within air-to-air and air-to-sea scenarios in all weather conditions. It is designed for installation on small business aircraft and has been sold to the Swedish air force and is in production for Brazil's SIVAM project for surveillance of the Amazon rain forests. Ericsson is marketing the ERIEYE system to other allied countries.
"ERIEYE's acceptance by the Greek air force is a strong indicator of both the power and economy of this airborne defensive tracking system," says Lennart Joelsson, general manager of Ericsson Microwave Systems. "Mercury's state-of-the-art products were selected after extensive evaluations of several alternatives, and their technical expertise for this program is unsurpassed."
"This first contract with a customer from a NATO country is a major breakthrough for Ericsson's ERIEYE system," says Andy Pine, managing director of Mercury Computer Systems Ltd. in England "Two export successes of ERIEYE within two years confirm the acceptance of this new generation system."
"In addition to the ERIEYE radar system, our relationship with Ericsson has resulted in the selection of Mercury technology for the multipurpose radar systems development for the SAAB Gripen, the Swedish Air Force's multimission fighter aircraft," Pine adds. - J.M.
Who's Who in DSP Boards
Catalina Research Inc., Colorado Springs, Colo., 719-637-0880, www.cri-dsp.com
CSPI, Billerica, Mass., 978-663-7598, www.cspi.com
Ixthos Inc., Leesburg, Va., 703-779-7800, www.ixthos.com
Mango DSP, Norwalk, Conn., 603-496-1284, www.mangocom.com
Mercury Computer Systems, Chelmsford, Mass., 978-256-1300, www.mc.com
Pentek Inc., Upper Saddle River, N.J., 201-818-5900, www.pentek.com
Radstone Technology PLC, Towcester, England, +44 (0) 1327 359444, www.radstone.com
Sky Computers Inc., Chelmsford, Mass., 978-250-1920, www.skycomputers.com
Spectrum Signal Processing, Burnaby, British Columbia, 604-421-5422, www.spectrumsignal.com
Synergy Microsystems, San Diego, Calif., 858-452-0020, www.synergymicro.com
Transtech DSP, Ithaca, N.Y., 607-257-8678, www.transtech-dsp.com
Mango uses MATLAB to improve embedded design
NORWALK, Conn. - Engineers at Mango Computers in Norwalk, Conn., and The MathWorks in Natick, Mass., are improving development times for military and aerospace engineers with a new tool that ports algorithms developed in MathWorks MATLAB language directly to embedded digital signal processor (DSP) configurations without converting to C code.
The Mango Math-Link EDS process eliminates the time-consuming task of hand-coding to shorten design cycles, reduces coding errors, and improves the quality of executable code, Mango officials say.
It cuts a Mango customer's development time by a factor of two, says Jerry Clancy, vice president of sales and marketing at Mango.
"We needed a way to shorten the development process for our acoustic signal processing applications and make our team more effective with its time," says Bill Knaack, technical director of Fiber Optic Acoustic Systems at Litton Guidance and Control in Woodland Hills, Calif.
"Now, with Mango's Math-Link EDS, we have an integrated, MATLAB-based real-time DSP environment, which enables us to eliminate entire phases from the development cycle," Knaack says.
The new tool provides a direct link from MATLAB to Analog Devices and Texas Instruments DSP chips in embedded multiprocessor configurations, enabling users to produce executable code from MATLAB M-files without leaving the MATLAB environment, Mango officials explain.Traditionally, engineers were forced to translate their M-files by hand to C or C++ and then enter into a lengthy development process in order to produce embeddable code.
"We believe the integration of The MathWorks and Mango's technologies offers an unbeatable combination for DSP developers," says Mike Berlin, president of Mango Computers. "With MATLAB as the front-end to our real-time multiprocessing development environment, Mango Math-Link EDS allows the designer to leapfrog directly to DSP code."
The Mango Math-Link EDS development system consists of a standard PC with MATLAB, the Mango compiler and run-time executive, and one or more of the company's multiprocessing DSP boards. Developers can easily edit, compile, debug, and download executable DSP code directly to any of Mango Computers' boards, without the need for additional hardware, software, or emulators, Mango officials claim. Its advanced debugging capabilities, such as trace, breakpoints, single-step, and plot, can be performed within the MATLAB interface.
Users can see how the M-code is performing in the target environment and can make necessary corrections quickly and easily, they claim.
"Mango Math-Link EDS allows MATLAB users to rapid prototype and deploy multiprocessor DSP applications, including real-time I/O data manipulations, without having to develop and hand code C or C++ programs," says Richard Drohan, Embedded Systems marketing manager for The MathWorks. "This greatly speeds product development and ensures a higher quality product. By maintaining DSP designs in MATLAB, developers are more efficient and can easily iterate their designs without needing to know the intricacies of the target DSP development tools or hardware." - J.M.