The challenge of moving from DSP boards to DSP subsystems
Once the digital electronics industry featured a plethora of board manufacturers that specialized in single-board digital signal processors.
By J.R. Wilson
Once the digital electronics industry featured a plethora of board manufacturers that specialized in single-board digital signal processors. Today, these same board manufacturers are re-inventing themselves into subsystem designers and systems integrators, which confronts them with a broad range of cooling, power, and software issues.
The DY 4 ChampAV conduction-cooled quad G4 DSP VME board is among the latest offerings of single-board digital single processors. At left is the DY 4 Reconfigurable Input Output Interface PCI mezzanine card.
Rapidly advancing technology is bringing new potential and new problems for digital signal processing (DSP) in military and aerospace applications. This progression is also redefining the way in which many board makers are serving their customers.
One major trend has vendors moving away from providing only boards or kits, and toward building complete subsystems or systems.
"We see that as the wave of the future, with fewer customers trying to build their own systems by buying boards and more looking for complete systems — hardware and software," says Bob Hoenig, chief technology officer for Sky Computers Inc. in Chelmsford, Mass. "Customers are less interested in the specific technology. Historically, they would come and ask about the board or card; now the movement is toward customers being more applications oriented — how many synthetic aperture radar images can you process per second rather than the exact technology inside the box? That gives us more flexibility to choose a custom ASIC [application-specific integrated circuit] than in the past."
Richard Jaenicke, marketing director for Mercury Computer Systems in Chelmsford, Mass., says the ideal is to provide customers with an overall system "where they don't need to know what boards are in there, only what the interfaces are to the outside world and how well it performs on the inside." However, he adds, many military DSP customers have custom applications for their boards and still want to build their own systems, with only design help from the board makers.
"Many of our customers still think in terms of boards and we need to address them even as we describe to them the benefits of thinking about systems instead of boards," Jaenicke says. "It really has to do with the ownership of the integration effort and the level of detail you need to know to put together the system. Many of our customers want to rely on us to know the very low level of detail needed to do DSP system integration. They've already decided their core competencies are at a higher level of incorporating DSP subsystems into overall platforms or systems within platforms, such as radars."
Sky Computers's marketing and sales vice president, Ed Hennessy, says that kind of effort has changed his company's whole role from the days when his engineers were primarily board suppliers: "Today we are drawn into sessions with defense contractors and the military to define the architectural design and systems capabilities," he says.
"Historically, DSP has been an auxiliary function; now it is a complete system," Hennessy continues. "There are key themes here — the evolution of the application development environment tools and the whole key area of middleware. We talk about simplicity and ease of programming, what we call QAI — quick application implementation. This is where you provide all the environment tools to develop an application quickly. We're even extending ourselves now to the development of applets."
Many DSP providers also are moving toward the fourth-generation (G-4) PowerPC microprocessor — the 7400/7410 series — with AltiVec technology, Motorola's high-performance vector parallel processing expansion to the PowerPC RISC processor architecture. AltiVec adds a 128-bit vector execution unit operating in concert with the PowerPC's existing integer and floating-point units to provide highly parallel operations, up to 16 simultaneously in a single clock cycle.
"AltiVec has made the PowerPC superior at DSP rather than just on a par with others," says Frank Phelan, principle engineer at Synergy Microsystems in San Diego, Calif. "The basic times we got with PowerPC before incorporating AltiVec wasn't a clear win against other parts, although it had an advantage in being a general purpose processor as well. With AltiVec on the G-4 7400/7410, we're seeing factors two to four times faster than the [previous device], the G-3 750. AltiVec's trick is it does four things at once, with an instruction set designed to take advantage of numerically intensive DSP algorithms."
Because military applications for DSP are a niche market, most companies leverage that line off commercial products, although military requirements — especially for aerospace applications — are far more demanding. To meet those, DSP manufacturers militarize their commercial products, with an emphasis on retaining — or even enhancing — functionality and performance.
"Building rugged versions of our DSP hardware is getting easier to do with plastic packages rather than the old ceramics," Phelan says. "The plastic packaging makes it easier to meet the temperature extremes, which means users can buy a fairly inexpensive commercial product to do their development, then not invest in the rugged product until they are ready to ship to their customers. Architecturally, our rugged line is identical to the commercial line, but we will make rugged versions up to and including full-conduction boards for the military. The electrical design and software are identical, so we can amortize those costs across both boards. The only added cost is to handle vibration and conduction cooling."
Power and cooling
Power and cooling are major problems for DSP board and system designers, especially when dealing with the relatively high requirements of the military and the rapidly advancing speed of chips.
"In systems that require more than 100 processors — which is our average system size — we have to very closely design or help our customers design the power and cooling systems," Jaenicke says. "We have not only standard VME and PCI chasses that have power and cooling systems with greatly enhanced capabilities, but we also provide a level beyond that — a system designed from the ground up as a large multiprocessor system with some standard VME bus slots but with the majority of slots being customer slots designed specifically for the highest power and cooling requirements of the military environment.
"We see power and cooling becoming more and more of an issue in the future," Jaenicke says. "And it is somewhat driven by the fact that people are starting to use more DSP-enhanced RISC processors rather than traditional DSP processing chips for high-end signal processing, such as radar and sonar. The PowerPC chips with AltiVec technology have almost universally been adopted for those applications."
Keeping up with cooling requirements can be one of the most difficult areas in board design, especially as ever-larger chips push the envelope for heat targets specified by the different bus architectures, such as VME and PCI.
"With the frequency of the chips going in excess of gigahertz, there is a huge concern over this almost exponential growth in heat dissipation," notes Gord Finlay, technical analyst at Spectrum Signal Processing Inc. of Burnaby, British Columbia. "Most chip manufacturers are working hard to figure out how to maintain the Moore's Law curve without an explosion of heat dissipation. Many VME board level products used within a chassis exceed the specifications per slot, which means you have to have an extra powerful fan inside the chassis to compensate for not being able to maintain the power spec."
Synergy's Phelan says the problem is especially acute for customers trying to plug new-generation boards into old backplanes and chasses designed around VME or Compact PCI standards.
"People are forever buying new VME64x boards with five-row connectors and trying to plug them into legacy chasses with only three-row connectors," Phelan says. "The connector will plug into a three-row backplane, but then you are trying to draw more power through six pins instead of eight or nine, which will work, but you don't have the right cooling design because they were used to 30-watt boards, not 50-watt, and you also could melt a connector, which just aggravates the problem. One good thing about the PowerPC in this regard is it has thermal management built into it so it will tell you how hot it is."
Jaenicke says another advantage to the PowerPC is its ability to draw less power than a typical desktop reduced instruction set computer (RISC) chip, which makes it a good prospect for embedded processing applications.
"But even PowerPC is not immune to the power per chip increases we've seen across the board for RISC processors," Jaenicke warns. "While the gigaflops per watt is going down, absolute watts per chip is still going up, which is a major design issue for buyers of board and system level DSP solutions."
Randy Tkatch, director of research and development for Spectrum Signal's Wireless Systems Group, says systems-integration expertise — designing an entire subsystem rather than just providing boards — is an advantage integration experts can carefully control the heat and power parameters throughout.
"One interesting thing about power dissipation these days is it can vary according to the software being used," he adds. "For instance, the actual algorithms being processed can greatly vary the power dissipation in the chips, so we also have to take that into account when developing the system."
Phelan says meeting the wide range of voltage requirements found on a single board has led to a maze of circuitry — and each new generation of chip just compounds the problem.
The DY 4 ChampAV conduction-cooled quad G4 DSP VME board follows today's trend of placing multiple DSP chips on one printed circuit board.
"Back in the 68k days," he says with reference to the old Motorola 68000 general-purpose microprocessor, "you had 5-volt power supplies, then 3.3 came around, then PowerPC came along at 1.2 volts. Now there are four or five different power supply voltages on our boards. We just keep adding more and more little power supply circuits on board to regulate down to what each chip wants — and each chip seems to want its own voltage. It's driving us crazy. But if you want the speed, you have to run them at the voltages they demand, he says.
"We also specialize in building symmetric multiprocessing boards," Phelan continues. "That became possible when the backside caches were introduced on the PowerPC 750. So running four of these at once, each with four megabytes of backside cache, enables all four to process data at the same time without getting into memory contention issues on the global memory. But these are power-hungry processors, so you have a huge power management headache."
Synergy Microsystems officials offer their own programmable power modules to meet the need to change voltages, but when a chip is added that requires an additional voltage, another supply has to be added. And, says Phelan, the chip manufacturers have made it clear users can expect power voltages to change with each new chip.
"Boxes have the problem, too, but not to the extent the chips do," Phelan points out. "As we get faster and faster chips, they reduce the core voltage. So as a board manufacturer, we're stuck with taking what comes in off the backplane and regulating to the voltage the parts need."
While the PowerPC is becoming increasingly popular in DSP board and system designs, other architectures remain an important part of the effort.
Alternative DSP approaches
"Hybrid architectures also are a big part of our future — a blend of high performance FPGAs [field-programmable gate arrays] and programmable DSPs," says Finlay. "The performance trend of DSPs is such that in order to address the requirements of future waveforms, we really need a hybrid architecture using FPGA and DSP technology, as appropriate."
Hoenig says while that does run counter to the idea of standards, adding some functional units to a core and teaching the compiler how to use them still enables writing in the higher language. He predicts that will become more and more common in the next three to five years.
"The first step to that is to configure a processor when the board is designed or the application is decided upon," Hoenig says. "Eventually, maybe five years out, we will be able to reconfigure the processor on the fly, which would be the ultimate in efficiency. All of those devices would be based on a standard PowerPC core, but with additional application units that will be configured when it is designed for a particular application and later on, possibly, on the fly. The trick isn't the hardware, but the software. You need an extensible compiler, otherwise the concept doesn't have any use."
The hybrid approach also can be related to relatively old systems, says Ian Stalker, DSP military products manager for DY 4 Systems in Kanata, Ontario. Experts from DY 4 and that company's wholly owned subsidiary, Ixthos Inc. in Leesburg, Va., are putting a high density FPGA and memory on the PMC-GPIO (general purpose I/O) card.
"In many of our military systems, there are legacy I/O devices that don't adhere to any particular standard," Stalker says. "They can buy COTS signal processing hardware, but they have to figure out how to get the data into it from these legacy systems and are interested in using FPGAs for that. There is a lot of capability within a modern programmable device. So I would expect to see a wave of clever applications of FPGA — not as the core of the solution, but as an accelerator at the front or back of the problem."
Spectrum Signal's Tkatch says the high horsepower and configurability of such combinations is especially important for developing real-time reconfigurable software-defined military radios.
Designers generally accept three forms of reconfigurability:
- static, in which the designer modifies a digital radio system air interface by uploading a new waveform, which would be relatively infrequent;
- pseudo-static, where designers modify or reconfigure the system between calls in the field to handle a new interface, such as spread spectrum; and
- dynamic, where designers change the modulation scheme during an established session over the air.
Tkatch says PowerPCs and other DSPs can support all three modes, while FPGAs can change the mode in a matter of microseconds, but ASICs cannot because they are not dynamically programmable.
"We see this being applied to three basic areas: Spectrum monitoring, up to hundreds of simultaneous channels; milcom, especially two-way air communications; wideband analysis systems," he says. "There is a huge potential market for tactical radio systems than can adapt."
Stalker says he agrees. "Implementing radios with DSP is a major growth area for us," he says. "We're doing a lot of work with some I/O partners to integrate the digital radio high-end receivers and downconverters. The military has much more demanding and sophisticated radios and this is a new area for them as well as for us, using a combination of radio receiver front ends with our DSP solutions. There is more of a crossover than average there between what goes on in the military and commercial worlds because a lot of telecom people also are interested in digital radio and getting a lot of channels into a small space."
Connectivity is another major issue for board designers, albeit one that is somewhat alleviated if they also handle subsystem configuration.
"There is a lot of discussion about which interconnect will win the war, with everyone striving for standard rather than proprietary connections," Sky's Hennessy notes. "A lot of that discussion revolves around InfiniBand and RapidIO."
There are significant disagreements over which of those — if either — is the best answer for future DSP applications (see sidebar page XX), an argument that has been in progress for at least the past two years as standards for both were being developed.
But there are other issues most DSP customers find far more pressing than the InfiniBand/RapidIO debate — and high on that list is the ability of a subsystem or system to keep up with ever-evolving technology and requirements.
"There is an underlying technology called vectorizing compiler technology," Hennessy explains. "We can demonstrate not only a front-end ease of programming, but when they are ready to migrate to the next generation — call it technology insertion — they can transition all their codes and get the same level of optimization with the next generation. That is critical to lifetime support."
Support is always a crucial issue, many agree. "Gone are the days when you could give a customer a manual on the DSP chip and he's off and running," Hoenig says. "Today, generation-to-generation compatibility is a major issue. It's a systems solution issue, which includes hardware, software, mechanical design, cooling, power, maintenance, technology insertion, etc. The bottom line is DSP is growing up in the software development area."
Jaenicke says he customers are placing a high value on programming productivity when it comes to DSP boards and subsystems.
"RISC processors are traditionally easier to program than DSPs and have a greater range of tools," he says. "With PowerPC being broadly used in the embedded market, there are many tools available, so it is possible to apply mainstream programming tools down into DSP applications."
One example, Jaenicke adds, involves Green Hills Software in Santa Barbara, Calif., which produces the MULTI 2000 Integrated Development Environment, which Mercury experts use to help program their DSP systems. "Using the Green Hills product will greatly reduce the time to deployment and the cost of development for system integrators," he claims.
A tale of two I/O architectures: InfiniBand vs. RapidIO
As demands on digital signal processing ramp up, so will the demands on system bandwidth. New and old technologies are vying for communications honors to meet that challenge and DSP board and subsystem producers are dividing into strongly opinionated camps.
At the heart of the debate are two new approaches — InfiniBand and RapidIO — with two companies based in Chelmsford, Mass., taking diametrically opposed positions. Sky Computers Inc. leading the charge for InfiniBand, while Mercury Computer Systems rallies the flag for RapidIO.
"RapidIO can provide from one to eight gigabytes per second of communication bandwidth on every com link. At that point, you will have a very good balance. We need that in the future for when the processors run at even higher speeds to balance the computing and communications," argues Mercury marketing director Richard Jaenicke.
"InfiniBand is designed to go between systems, which we love it for. It is perfect for communicating between servers and clusters," Jaenicke says. "It can be applied to subsystems communication in military systems, as well, but it definitely is not designed for use inside a system, where processors need to be tightly coupled to perform a single function. RapidIO was specifically designed for that."
Sky Computers chief technical officer Bob Hoenig takes the opposite view.
"We have taken a position in favor of InfiniBand; we believe it has the much larger market support out in the field and there will be more parts available," Hoenig says. "The second point is technological — InfiniBand is a serial implementation and RapidIO is parallel. Serial can do optical interconnects, which would require a redesign with RapidIO. There is some overlap between the two, however, and we believe you can use InfiniBand inside the box as easily as you can use RapidIO."
Jaenicke, however, points out that a serial version of RapidIO will emerge soon to counter some of that argument, while offering a few negatives of his own about InfiniBand for internal communications applications.
"The current version of RapidIO is a parallel connection because it was meant to go between chips on a board. Inside a system the number of wires isn't so critical," Jaenicke insists. "That's not to say there's not some concern to keep those under control. RapidIO brings the number of onboard pins down to 40 from 100 or more, but there really isn't any need to go down to 4. Each InfiniBand link is four wires on a base serial link — one pair going in each direction.
"Typically those are grouped together in four — or 16 wires," Jaenicke continues. "The number of wires in a connection is very important when going between systems, which is what InfiniBand was designed for. Serial RapidIO is still in definition, but a logical guess is it will have two differential pairs, giving you the option of which underlying physical layer you want to use while still using the RapidIO protocol that was optimized for communication inside a system."
Hoenig, in turn, disagrees on the wires issue: "The reduction in the number of wires is important — 16 per link for InfiniBand versus 76 for RapidIO. It makes a major difference. There also are some practical, physical implementation issues."
Jaenicke says it also is possible to connect systems with serial RapidIO or even use parallel RapidIO for short distances, such as two chasses in a rack or two racks next to each other, but says he does not recommend it.
"RapidIO doesn't have a software protocol stack — it is entirely implemented in hardware meant for trusted transactions between peers in a network, where one processor is allowed to write into another processor's memory," he says. "You want that inside a system because it allows very fast communication with very low latency. That's not exactly what you want to happen between systems, where you don't want one system overwriting the memory in another system. InfiniBand is entirely a message passing-based protocol, with a heavy protocol stack to protect one system from the others connected to it. So it is ideal for connecting between systems, but is a much heavier protocol than what you want for communication inside a system."
RapidIO also has won the endorsement of Spectrum Signal Processing Inc. in Burnaby, British Columbia.
"We see parallel RapidIO being important for intraboard communications, but over the backplane within the chassis, high performance gigabit serial will be important," says Spectrum Signal technical analyst Gord Finlay. "That probably will be serial RapidIO, but there are a number of competing technologies."
Randy Tkatch, director of research and development for Spectrum's Wireless Systems Group, says their evaluations of the two technologies have them planning to adopt serial RapidIO as part of their future architecture, but using InfiniBand for interchassis communications outside the box. And they expect the internal configuration to be simplified by Motorola and other chipmakers embedding RapidIO ports, probably one or two 8-bit devices, in future processors.
Others in the industry, however, are less enthusiastic about either — or about the prospects for embedded RapidIO.
"I think you're at least a year away from Motorola doing a RapidIO interface — and that won't happen until Apple decides they want it," says Frank Phelan, principle engineer at Synergy Microsystems in San Diego, Calif. "And does Apple really require that much extra bandwidth?" he asks.
"The DSP world just doesn't have enough size to force that kind of decision, Phelan continues. "And it's only the cutting edge stuff that needs that much extra performance even there and you have only a limited number of people willing to pay for it. If there are chips available with a RapidIO interface, then we'll probably build boards with RapidIO. But what we can get now are PCI-based. There are hundreds of cards you can plug into our boards — all PCI-based. This debate has been going on for a couple of years now, but there still are no products you can build with it."
Ian Stalker, DSP military products manager for DY 4 Systems in Kanata, Ontario, and their wholly owned subsidiary, Ixthos Inc. in Leesburg, Va., says he agrees.
"It's not all that relevant for what people need to do today, which is build DSP systems with elements available off the shelf today," Stalker says. "Clearly the world is going to switched fabric mechanisms, which RapidIO and InfiniBand are. We're staying in touch with what technology will be available to us and when — and when the time is appropriate, we'll build products with those technologies."
For example, in January Ixthos leaders announced their new PMC GPIO-1 user-programmable general-purpose card, which enables designers to configure the module to meet a wide range of I/O needs for a variety of applications, including a RapidIO implementation.
The new card provides 64 user-defined, bi-directional, single-ended or differential I/O signals, each with the option of driving or receiving LVDS, LVTTL, CMOS, and other levels, with on-board termination. With a high performance 64 bit wide, 66 MHz PCI interface to the baseboard, up to 256 megabytes of on-board SDRAM memory and incorporating direct memory access and interrupt capabilities, Ixthos officials say the PMC GPIO-1 is ideal for implementing custom serial or parallel I/O protocols, including RapidIO with or without data buffering.
In the meantime, DY 4 engineers also are working with other technologies that are available now.
"We play in the extreme rugged end of the military market — helicopters, fast jets, IR systems in tanks and things of that nature," Stalker says. "In the past year or two, we've been designing a lot of Fibre Channel products into those applications because we have our own Fibre Channel technology, as well. We're one of the few who can put Fibre Channel into a conduction-cooled ATR box."
So far as InfiniBand versus RapidIO is concerned, much of the industry is adopting Phelan's "wait and see" attitude.
"We'll design what wins into our products, so long as the performance is there," Phelan says. "But other companies need to pick up those standards and deliver the port logic — like gigabit Ethernet to InfiniBand or RapidIO to Fibre Channel."
Until there is a wealth of peripheral support for these, all you can do is make one processor talk to another," Phelan concludes. "Until I start seeing processors and peripherals with those interfaces, its just an argument that really doesn't mean anything." — J.R.W.