DSP Design: Faster, Better, Cheaper – and Smaller

by James Henderson, President, Innovative Integration Inc.

Although it’s been said the credo of “faster, better, cheaper” is impossible, the cold, hard fact is that the high-pressure world of embedded DSP design demands all three. Not only that, this brave new world also demands embedded devices be smaller, tougher – and first to market.

Overview

Time-to-market is critical. DSP designers can’t afford to waste time and money on NRE costs when customizable COTS devices are readily available. Flexible high-resolution waveform generation, digitization and analysis subsystems capable of manipulating RF signals in conjunction with digital down-conversion (DDC), fast Fournier transforms (FFTs), and tuning multiple regions of interest are required. Subsequently, real-time, multi-channel demodulation of these regions using a wide variety of schemes is required. Often, such equipment must be portable and operate under harsh environmental conditions creating serious challenges in packaging and power consumption.

Existing solutions employ arrays of dedicated DSPs working in tandem with an RF digitizer to provide the computational bandwidth needed to implement DDC and demodulation functions. Though effective, this approach is complicated and expensive, because multiprocessor programming requires sophisticated process management and load balancing. At the same time, race conditions and data bottlenecks must be avoided.

New solutions are emerging that use highly modular devices leveraging industry standards. These solutions include (computer-on-module) COM-Express PC architecture, advanced software development tools, and PCI Express (PCIe) -based XMC mezzanine modules to create cost-effective, customizable RF-processing blocks.

Cutting PCs down to size

The performance of mainstream DSP devices has effectively stalled. The system clocks on such devices are currently 1 GHz or less, and bandwidth throughput is limited to about 800 MB/s on the common 100 MHz, 64-bit external buses. Meanwhile, Intel and AMD continue to improve the x86 architecture by adding instruction set optimizations, enhanced caches, floating point co-processing, and multiple cores on chips. Moore’s Law still applies. Quad-core processors featuring 3 GHz processing CPUs and 5 GB/s external bus bandwidth are ubiquitous.

Moreover, the superb Intel Performance Primitives support native signal processing on an x86 processor that is an order of magnitude faster than existing DSP devices. Add to this the convenience and accuracy of 80-bit floating-point capabilities. Nevertheless, desktop and industrial PCs do not meet the portability, packaging, or environmental extremities necessary for many embedded applications.

Fortunately, the immense popularity of the PC as a development and processing tool has made the COM-Express format the de-facto standard amongst users requiring the utmost reliability, scalability, portability, and computational performance.

COM-Express modules are commodity items available from a number of reputable electronics vendors such as Kontron, Radisys, DTI, and Advantech. Pricing ranges from approximately $400 to over $2000/unit in unit quantities. Pricing is primarily a function of computational capabilities and the required temperature, shock, and environmental capabilities required. COM-Express modules are small, mezzanine modules that are mounted onto carrier boards that are customized to meet application-specific requirements. Figure 1 shows a typical COM-Express card.

Figure 1: COM-Express is a rugged PC embedded on a standard-sized small circuit board.

While the cost and computational advantages of a COM-Express PC compared to traditional chip-level DSP solutions are incentive enough, an additional benefit of the COM-Express architecture is the ability to leverage the existing body of excellent development and debugging tools available for the PC. Whereas Texas Instruments or Analog Devices are sole sources for compiler and debugger tools for their DSP devices, the PC marketplace boasts thousands of well-established vendors that offer sophisticated, mature software featuring superior performance at a reduced cost. Even so, use of COM-Express configurations in the final product does not preclude the use of conventional desktop PCs to accelerate and simplify the development of COM-Express-based products.

Expanding I/O and DSP capabilities

Though powerful computationally, a COM-Express PC does not directly support the acquisition or analysis of RF analog signals. Moreover, the multi-core x86 CPUs available now and in the foreseeable future do not have the bandwidth necessary to process RF signals in real time. Therefore, some form of I/O and processing expansion is required.

Fortunately, COTS I/O cards are also available in small form factors ideal for use in embedded instrumentation. PMC (PCI Mezzanine Card) modules support both PCI and PCIe communications. PCIe is the heir-apparent to the older, slower PCI bus. Technically, PCIe is not a traditional bus but rather is comprised of point-to-point full duplex serial links called lanes. The primary advantage of PCIe technology is its dramatically enhanced throughput, which is up to 64 times faster than a PCI bus. Additionally, PCIe cards guarantee QoS (quality-of-service) data flow capabilities, making it an excellent choice for real-time applications. PCIe modules are essentially identical to those used in standard desktop PCs, except they – like COM-Express cards – are packaged in small, rugged formats. And just as PCI-Express has replaced PCI as the technology of choice, PMC boards are being replaced by newer XMC modules.

The eInstrument solution

To take advantage of these technological advances, Innovative Integration has created the highly customizable eInstrument. Featuring a special carrier board, the eInstrument accommodates a rugged, Intel-based COM-Express PC board, sites for two PMC/XMC expansion modules, and an array of integrated peripherals. Designed for high-powered versatility in a small form factor package, the eInstrument can be configured and embedded within OEM equipment to create intelligent, autonomous instrumentation, real-time servo control, or RF processing nodes. Figure 2 shows a typically configured eInstrument.

Figure 2: This ruggedized eInstrument has an embedded COM-Express PC with two XMCs packaged into an 8” x 8” x 2.5” (20.32 cm x 20.32 cm x 6.35 cm) chassis 1U x 1/4 rack width).

eInstrument devices eschew legacy PMC boards in favor of the higher-bandwidth PCI Express XMC modules, such as Innovative’s own family of XMC mezzanine cards. Any mix of the standard peripherals typically found on a PC can be made available in an eInstrument-based system – including Ethernet communications, disk drives, USB, and SATA ports. Keyboard and video ports can also be provided to aid in field diagnostics. Thus, it is entirely feasible that an eInstrument assembly located in a remote site – or in orbit – could be accessed via the Internet or VNC to provide interactive support or software upgrades.

The eInstrument COM-Express PC performs initialization, supervisory control, and user interface as well as high-performance computational duties in RF processing applications. The configuration of a COM-Express PC with two PMC/XMC modules for I/O and DSP expansion is illustrated in Figure 3.

Figure 3: This block diagram illustrates how a COM-Express PC interfaces with the two PMC/XMC sites on an eInstrument carrier board.

Diagnostic ports for keyboards, mice, buttons, various displays, USB and hard disk interfaces may be exposed, hidden, or omitted in the initial design or presented as optional devices available only on designated systems.

Two XMC module sites are provided for I/O expansion. One is typically used to host a module that implements the RF front-end analog I/O and FPGA-based digital signal processing capabilities. The second site is uncommitted and available for future expansion. A bevy of PCIe-compliant XMC modules are available on the market to provide additional capabilities such as Fibre Channel or Ethernet communications, auxiliary voice or ultrasonic-band analog channels, or additional FPGA resources.

Each eInstrument site features four or eight 2.5 Gbps PCIe I/O lanes, which are essential to support sustained high-speed data transfers. Additionally, the two sites provide eight dedicated communications lanes to allow implementation of algorithms in which large volumes of data are shared between modules. Even accounting for lane inefficiencies, sustained inter-site data rates of 1.2 GB/s can be realized.

The PCIe lanes provide excellent, high-bandwidth connectivity between the COM-Express CPU and the PMC/XMC modules. Given the 1 GB/s sustained throughput, the interface has sufficient bandwidth for low-bandwidth DDC baseband data plus plenty of additional bandwidth in reserve should it become desirable to capture or log raw IF (intermediate frequency) data in future applications.

PCIe  also supports fast, random, asynchronous I/O accesses to peripheral registers on XMC modules to accommodate operations such as filter coefficient uploads, DDC channel tuning, and the myriad other operations typically required in software radio applications. Individual slave-type accesses will complete in under 1 µS in a COM-Express module.

Xilinx FPGAs on board

Innovative's two new X5 modules combine the most powerful programmable logic devices ever offered by Xilinx – the Virtex-5 family of FPGAs. These modules combine up to four channels of high-resolution analog I/O plus FPGA-based signal processing cores capable of performing the signal digitizing, data buffering, and signal processing required for real-time RF processing applications. An X5-210m module is pictured in Figure 4.

Figure 4: A Xilinx Virtex-5 FPGA is the core component of the X5-210 XMC module.

These XMC modules support conduction-cooled operation in accordance with the VITA 20 mechanical specification. Additionally, the standard logic includes continuous temperature monitoring. Whenever the temperature exceeds a programmed warning threshold, the logic automatically shuts down the device. This advanced thermal management insures excellent in-field reliability.

The COM-Express PCs embedded in eInstruments run standard Windows or Linux variants such as XP or openSUSE to fully utilize drivers available for existing PMC/XMC modules. These PCs are provided with C++ libraries, which exploit high-performance signal processing features in the optimized Intel Performance Primitives library, yielding world-class DSP functionality and performance running on any standard x86 platform.

Custom firmware for the FPGAs may be built using standard IP cores and fully modeled under MATLAB, which facilitates high performance and accelerated time-to-market for embedded applications. Custom firmware for Virtex-5 FPGAs builds on Innovative’s FrameWork Logic software to interact with the on-board analog devices, DDR and QDR memory pools, and PCIe lane interface. The firmware works with in conjunction with PC-based software tools and C++ libraries, providing a comprehensive software development system for integration of the PMC/XMC with the host application.

To provide optimal AC performance, high-speed RF analog input circuitry is driven by a stable, low-jitter sample clock. The onboard clock circuit exhibits <1 pS RMS jitter for a 6.25 MHz to 270 MHz clock range. Additionally, long-term thermal stability and integrated clock drivers are capable of simultaneously sourcing into 50W loads on each of the XMC sites and external devices through the EXT CLK connector.

In some applications, control logic embedded in the carrier FPGA servo-locks to the epoch (1 pps) output events produced by a GPS receiver, insuring that eInstrument PCs located in disparate locations around the world start acquisition and sample synchronously to within 1 µS.

Embedded signal processing

eInstrument X5 modules are engineered to support RF signal processing applications with minimal external circuitry and with no modification of the PCIe back-end infrastructure. For example, Figure 5 shows a functional block diagram of the X5-210m XMC module.

Figure 5: This block diagram of an X5-210m features a state-of-the-art Xilinx Virtex-5 FPGA with a high-performance DSP core and a PCIe interface for system integration.

As with both X5 (210m and 400m) modules, FPGAs can be programmed using either HDL or MATLAB using Xilinx System Generator. Typically, the FPGAs are modified to implement independent DDC channels, filters, FFTs, and other operations that must be performed at IF frequencies for baseline RF tuning functionality. The COM-Express CPU performs initialization, supervisory control, and user interface, as well as high-performance RF computational duties.

 The MATLAB board support package for the X5 modules allows signal processing to be developed using MATLAB/SimuLink development tools. SimuLink models signal processing for bit-true, cycle-true design that enables direct hardware-in-the-loop testing. This allows the signal processing to be developed at a high level, using proven Xilinx IP cores and testing in the MATLAB environment.

This simultaneous HW/SW testing technique reduces risk and shortens development time by allowing efficient and thorough verification of DSP from within the powerful MATLAB/SimuLink environment. The signal processing logic from MATLAB is then integrated into the FrameWork HDL for the final logic design.

The Framework Logic package that comes with the X5 modules provides the hardware interface and support functions, including A/D interface, memory controllers, host data interface, and controls. All standard logic features such as A/D interface, triggering, multi-queue data buffering, DDC control, and PCI controller interface are supplied as components that must be augmented with custom logic blocks. These are usable in either SimuLink or Xilinx ISE to form the foundation for the user’s application firmware.

Ordinarily, the desired, application-specific signal processing functions can be provided by Innovative Integration or another module manufacturer or engineering firms specializing in development of IP, such as RF Engines Inc.

To minimize NRE costs and maximize time-to-market, application-specific eInstruments can be co-developed by the Innovative Integration team and the client engineering staff. Of course, in cases where trade secrets or top-secret security is involved, Innovative Integration engineers can provide the training the customer needs to "hit the ground running" and finish proprietary development on its own terms.

Summary

In this brave new world of ever-evolving technology, embedded designers need all the help they can get to create devices that are faster, better, cheaper – and smaller and tougher. COTS small form factor COM-Express PCs and PCIe XMC modules are essential in the race to be first to market.

E-mail: jhenderson@innovative-dsp.com
Website: www.innovative-dsp.com