home contact us site map

A New Approach to Embedded Hardware

Jeffry Milrod
President, BittWare, Inc.

About a decade ago, the world of 'embedded' DSP underwent a dramatic change. For the first time ever, you could build complete systems with very little or no custom hardware. The emergence of commercial-off-the-shelf (COTS) DSP boards on one of several standard bus formats, STD, VME, and PCI, empowered this approach to system design.

At first DSP board vendors (a new concept then) released very simple, main stream COTS products that systems designers could use to jump-start their design efforts and reduce non-recurring engineering (NRE) by not having to 'reinvent the wheel.' As the COTS 'bus & board' method of system design caught on, DSP boards became more flexible and complex. System designers now have a broad selection of DSPs, architecture, memory, and even modular I/O available from a variety of COTS vendors.

A New Approach Needed

But here's the rub - the COTS 'bus & board' approach to system design provides ideal solutions to only a very limited segment of the embedded DSP market. That portion is the descendant of the main frame, big system approach used by the military and high-end test, measurement, and instrumentation houses. In this world, 'embedded' used to mean anything that didn't require a separate, stand-alone computer to control it. Remember, in the old days computers were BIG, and real-time systems of any complexity were really big. So using COTS to fit a complete system of this magnitude into a 21-slot 6U VME rack is quite impressive.

Although the COTS 'bus & board' market continues to grow, in the age of cell phones, palm-tops, set-top boxes, and DSP controlled washing machines, it's hard to imagine a 6U card as being embedded, much less a rack. And size isn't the only measure of 'embeddedness': most racks (3U or 6U) cost more without any COTS boards than the entire electronic assemblies of today's truly embedded systems. Not coincidentally, another parameter of 'embeddedness' that has changed dramatically over the last decade or so is quantity.

The fact is that the nature of embedded DSP systems has changed tremendously over the last decade or so, and most COTS solutions haven't fundamentally changed at all. They're still fairly large (3U/6U), fairly expensive, and require a backplane. The one exception to this rule is PC/104-Plus, which is smaller, cheaper, and backplane-less; it also happens to be the fastest growing COTS segment (although still relatively small in absolute size).

One alternative methodology that has emerged to address the truly embedded DSP market is System-On-Chip (SOC) solutions. Unfortunately, this radical approach addresses an even smaller (though potential much more lucrative) segment of the embedded DSP market than COTS 'bus & boards' does. Even the most optimistic SOC projections hope to win only 5-10% of the potential 'embedded' systems design-ins. A lot of money is being thrown at this approach because that 5-10% might well account for more than 50% of the dollars, but what about the rest of the designs that don't have the volumes to justify the large NRE?

Without attempting to quantify it (I'll leave that to the market experts like the esteemed publisher of this fine publication), it's obvious that the bulk of embedded DSP design-ins are not being well served by value-added solutions vendors. The fact is, even with all the wonderful COTS 'bus & boards' and SOC solutions available, most embedded systems designers using DSP must either compromise their design to use available solutions, or they must go back to the old days of 'roll-your own' and 're-invent the wheel.' The cost of this reality to the average system designer is enormous - not just in direct NRE costs, but also in increased risk and time-to-market.

What embedded DSP needs is an alternative approach between the extremes of the COTS 'one-size-fits many' complexity with it's high recurring costs, and the highly customized elegance and correspondingly high initial costs of SOC.

Reusable, Modular Board-Level Design - System-On-Board (SOB)

One such approach could be called System-On-Board, with the unfortunate acronym of SOB. Analogous to SOC, SOB utilizes existing macro cells, interconnect technologies, and interface software created from COTS 'bus & board' intellectual property (IP) to create quick turn, tailored solutions - but integrated at the board-level rather than chip-level. Therefore, an SOB approach requires much lower initial investments than SOC and, since it is tailored to a specific system, should be significantly cheaper per system than COTS. In addition, an SOB solution greatly reduces risk and time-to-market when compared to 'reinventing the wheel.'

It must be made clear that SOB is not just a marketing spin of the traditional customizations and variants historically done by most COTS board vendors any more than an ASIC is an SOC. SOB is not simply a team of experts 'rolling their own.' Rather, like SOC, an SOB solution requires a dedicated team using a structured approach of reusable, modular board-level and software designs.

Each board building block, or macro if you will, must have been designed with reusability and modularity in mind. This is an engineering methodology that is common in software design (or at least should be) and chip-level design, but it is fairly rare in board-level hardware design. Sometimes this requires some additional overhead circuitry, sometimes not, but it always requires thinking through interface, test, and layout issues not usually thought about. Further, each signal net must be have well-documented electrical specifications such as maximum allowable trace length, variation, and impedance to facilitate repeatable re-routing.

Ideally, each board macro would be fully synthesize-able and simulate-able, with a completely abstracted modular interface; i.e. any board macro can seamlessly and transparently connect to any other board macro. In addition, each board macro should have corresponding modular low-level software such as kernels, loaders, and drivers.

This design approach is particularly well suited to DSP, since customers need horizontal enabling technologies that they can use as system 'building blocks'. Much like the ubiquitous LEGO toy does for kids, an embedded DSP vendor's job is to provide system/product engineers the pieces they need to creatively and innovatively implement their applications.

Clearly, as with SOC, this is a non-trivial effort that will require years to build up a large and broad set of reusable, modular SOB macros. However, BittWare has started to do just that by narrowly focusing it's design efforts on embedding Analog Devices' SHARCâ family of DSPs. For over a year now all of BittWare's designs, from both their Standard and Application-specific Product divisions, have been done with SOB in mind. Granted, it's not been easy or ideal, but it's a start.

BittWare's unique value-added to this market is not in knowing how to do telecomm better than its telecomm customers, or instrumentation better than its instrumentation customers. Rather, it is in leveraging the great wealth of DSP expertise and experience acquired over the past decade to build better and more flexible DSP 'building blocks' that can be easily implemented into SOB application-specific solutions.

SHARC DSP Implementations Using Reusable, Modular Board Design

BittWare's new COTS Hammerhead boards, based on the ADSP-21160 DSPs, uses a modular architecture that was implemented as several reusable board macros. While imposing SOB methodology on the developers somewhat delayed the initial product release, it has greatly facilitated the proliferation of design variants. Within just six months of the initial release, four different Hammerhead implementations have been done with remarkably little additional effort.

Macros used on the Hammerhead product include:

SHARC DSP Cluster (1-4 SHARCs)

8 Off-board Inter-Processor Communication Link Ports per cluster (100MB/sec each)
2 Off-board Serial Ports per cluster (100Mb/sec each)

SHARC-PCI Bridge (64-bit, 66MHz)

Interrupt/Flag Multiplexor
SDRAM Controller (up to 512MB)
PCI Mezzanine Interface (64-bit, 66MHz)

FLASH & UART
Multi-Voltage Power Regulation with Boot Sequencing

BittWare's Hammerhead-6U-cPCI Board

Central to the success of the Hammerhead's design methodology is BittWare's SharcâFINä ASIC. The primary function of this chip is to act as a bridge between the ADSP-21160s and PCI; however, during the development it became much more. Despite best efforts, the board macros did not always work together as seamlessly as hoped. Since the SharcFIN was implemented using QuickLogic's Embedded Standard Product (ESP) technology, its FPGA portion can be easily modified. Therefore, the SharcFIN was tweaked from board to board, becoming the glue that holds the macros together and the caulk that fills the gaps.

Although it too soon to tell for sure, its quite possible that most real-world (i.e. non-ideal) SOB implementations will require some similarly non-reusable glue and caulk. In practice, this seems a small price to pay for the potentially overwhelming benefits.

Prior to these efforts, the Shortfin family of DSP boards were the first products developed after BittWare began to use reusable, modular design methodologies. These products are based on the low-cost ADSP-21065L and also feature PCI, SDRAM, and off-board communications.

Embedded SHARC DSP Implemented Using SOB Methods

The use of reusable, modular design methodologies in the development of standard COTS product is more of a means than an end, however. The end is to equip BittWare's Application-specific Product Division (APD) with modular SOB IP it can rapidly cost-effectively implement as value-added solutions specific to the needs of various customers. While the Hammerhead macros are still being transitioned, BittWare's APD has already developed deployed several SOB solutions of their own.

Based on the ADSP-20165L core macro from the Shortfin product, APD has already delivered several tailored SOB solutions to embedded OEMs. Additional macros developed and implemented in these efforts have included:

High-Performance Stereo Audio CODEC
High-Performance Digital Stereo Audio Interface
Multi-channel (8) Audio Interface
Video Interface
Man-Machine Interface

LCD
Rotary Encoder

Real-Time Clock with Battery Back-up
Watchdog Timers
Low-Power Consumption Controller

As a result of the macro reuse, some of these DSP board-level solutions were delivered to the customer in less than 8 weeks from original specification. Most of these have been done for OEM system designs that either would have required in-house development efforts from already overburdened engineering resources, ground-up custom development by service vendors, or wouldn't have used DSP at all. So far, it appears that this SOB approach to embedded DSP implementation is effective.

As BittWare's collection of board macro IP continues to expand, so too does the flexibility and responsiveness of the embedded DSP solutions it can offer system designers. As all future products developed by BittWare will use SOB design, the Standard Product Division is driving to refine the methodology, and continuously developing additional macros as new COTS products are released. Macro functions that will be available soon include: