Designing an embedded system based on PC-industry building blocks is like dancing with the devil. The chips and subsystems, including add-in cards, hard-disk drives, optical drives, and power supplies, are low-cost and abundant, thanks to the high-volume-manufacturing efficiencies of the PC market. However, although you measure your design's anticipated production life span in years or even decades, the fickle fortunes and fast evolution of the PC industry drive rapid obsolescence of your raw materials. Design smartly, planning the ability to later upgrade, and you'll be able to nimbly sidestep any supply-chain potholes. Failure to plan for future substitutions and advancements, on the other hand, means you'll soon—and perhaps repeatedly—redesign.

The PC industry's rapidly spinning product treadmill will become abundantly obvious to you if you revisit 2004's two-part article series (reference 1, reference 2). Then-state-of-the-art high-end systems are now mainstream products or have even moved to bargain-basement closeout status. Some of those systems' constituent pieces, such as Rambus DRAM and RDRAM-cognizant core-logic chip sets, have disappeared from today's PC designs. And multicore x86 CPUs, which in early 2004 were placeholders on manufacturers' future product road maps, now take center stage.reference 3

The Intel Pentium M processor, which in early 2004 was only beginning to establish a beachhead in the mobile-computer market it now dominates, has today also become a popular CPU in single-board-computer designs. Its combination of high performance and low power consumption makes it a natural fit not only in laptops, but also in many embedded systems. With Core Duo (formerly known as Yonah) now in production on Intel's 65-nm process, dual-core capability is now part of the Pentium M stable. Core Duo chips for the embedded-system world are now in short supply because companies such as Apple and Dell are gobbling up as many wafers as Intel can fabricate. As a result, this hands-on project substitutes laptops for single-board computers, tweaking test conditions to as closely as possible mimic embedded-system configurations (Figure 1 and Table 1).

A near-infinite number of possible combinations of hardware, operating systems, and applications exist; no one study can hope to encompass even a small percentage of them. This project uses reference systems that target various sizes, weights, prices, and performance and power-consumption rates in the hope of providing a spectrum of data points that you can extrapolate to your design requirements. It also harnesses Windows-based benchmark suites; again, you can use them as starting points for follow-on evaluations using your operating system and applications. This project broadens the speed-centric focus of the earlier articles to encompass power-consumption measurements. Will Core Duo deliver the performance that Intel claims and, if so, at what power or other trade-offs? Does a dual-CPU configuration deliver a credible alternative approach, and when does it make sense to defer both of these emerging options in favor of a traditional, single-core, single-CPU setup?

SiSoftware's free Sandra Lite program enables you to determine much information about a Windows or Windows CE-based system; Professional, Engineer, and Enterprise versions of Sandra deliver additional tests and test-configuration options. The earlier articles used the 2004 variant of Sandra, and Table 2 reiterates the high-level data from those articles on the Spirit and Opportunity test platforms. The various CPUs' integer and floating-point performance results closely correlate with the relative performance you'd expect, given the CPUs' clock speeds, on-chip cache types and sizes, and other integrated features. With both cores enabled, the Dell Latitude D820 delivered twice the arithmetic- and multimedia-benchmark performance of the same system with one core disabled using a BIOS setting. Later, you'll see how the Latitude D820 performed in single- and dual-core configurations with real-life applications.

The memory-bandwidth tests also produced no big surprises; note that the Fujitsu Lifebook-P1510D outperforms its Opportunity and Dell Inspiron 700m "big brothers" here by virtue of its higher speed DDR SDRAM and more modern chip set. You might scratch your head in bewilderment when perusing the file-system benchmark results: Why did the two-year-old Spirit and Opportunity setups run rings around modern laptops? Look closely at Table 1, and you'll find part of your answer: The earlier mini-ITX boards hooked up to high-capacity, 3.5-in. hard-disk drives. With that qualifier in mind, you may still be a bit surprised at the underwhelming performance of the Dell Inspiron 700m's Fujitsu hard-disk drive. Like the Maxtor drives that Spirit and Opportunity use, it's a 5400-rpm unit, and it also has four times the onboard buffer RAM of the mini-ITX boards' Maxtor hard-disk drives.

Spirit and Opportunity are now disassembled, and their constituent pieces inhabit my garage. A busy travel schedule didn't provide enough time for reassembly, so I was unable to run the newer Sandra 2005 benchmark suite on them. However, I did run Sandra 2005 on the three laptops (Table 3). As with Sandra 2004, you'll likely find no major surprises once you correlate the results to the systems' capabilities. Still, it's interesting to see the influence of, for example, various hard-disk-drive-system interfaces, capacities, rotational speeds, and buffer sizes on the benchmark outcomes.

Test a stand-alone processor, and its results may blow you away. However, if you subsequently test a system containing that processor and on real-life code instead of carefully crafted synthetic benchmarks, you might be less—or maybe more—impressed. To assess the capabilities of Intel's Core Duo CPU in realistic operating scenarios, I installed and fired up BAPCo's (Business Applications Performance Corp's) SysMark 2004 SE benchmark suite (Table 4). SysMark 2004 SE is ideal for multicore- and multiprocessor-system configurations, because it simultaneously multitasks between multiple applications and because many of the applications it runs are multithreaded.

With a few notable exceptions, the Dell Latitude D820 delivered approximately 50% higher SysMark numbers with both processor cores enabled—not the twofold improvement that Sandra's synthetic benchmarks suggested but nonetheless better results than I predicted beforehand. I configured the Latitude D820 at its LCD's native 1920×1200-pixel resolution, along with a 32-bit color depth. Dell had earlier run the Latitude D820 with both cores enabled through SysMark 2004 SE with a 1024×768-pixel resolution setting and obtained Internet-content-creation, office-productivity, and overall results of, respectively, 288, 172, and 222.

A speedy system is still unacceptable if it too quickly drains the battery, or, if ac-powered, it throws off excessive heat. To measure power consumption in various system operating modes, I picked up P3 International's $30 Kill A Watt and plugged the various laptop PCs into it (Figure 2). To simulate a no-display single-board-computer configuration, I turned off the systems' LCD backlights for all measurements. This approach made a notable difference in the results: The Latitude D820's screen, for example, single-handedly pulled an incremental 0.1A of current with the backlight fully illuminated. I also removed the systems' batteries before running my tests, so that incremental charging current, for example, wouldn't distort the results, and I disabled the systems' various network interfaces.

Before measuring idle-mode power consumption, I terminated all unnecessary background-running processes and waited for the systems' hard-disk drives to spin down and park. The Inspiron D820 results were erratic, probably as a result of the processor cores' moving through various operating states in response to incoming operating-system requests. So, I selected the lowest numbers I observed over a several-minute-long time interval. Conversely, in full-active mode, I ran Sandra 2005's cache and memory benchmark, which fully uses all available CPU resources and activates the hard-disk drive, in a repeating loop using the program's burn-in wizard.

Table 5 shows the differences between the Fujitsu Lifebook-P1510D, Dell Inspiron 700m, and Dell Latitude D820, which are respectively thin-and-light, mainstream, and desktop-replacement systems. Note, for example, the approximately four-times-greater incremental power drain of the Latitude D820 than that of the Lifebook-P1510D in standby mode. Idle-mode and full-active-mode comparisons between the two systems also reveal a two- to four-times increase in power consumption for the Core Duo-equipped system, which counterbalances its performance advantages.

However, the apparent power consumption of the Inspiron 700m in active mode is higher than that of the Dell Latitude D820. More generally, when comparing systems, make sure to factor in not only the power a system consumes when executing an operation, but also the time it takes to complete that operation. In other words, focus both on power and energy, which is a product of power times time.

Compare the Latitude D820's single- and dual-core configurations, and you find additional interesting data. In idle mode, the dual-core variant pulled approximately two times the apparent power of its single-core counterpart; this result correlates, in part, to the louder fan noise from the Latitude D820 at idle in dual-core mode. However, in full-active mode, the dual-core Latitude D820 consumed only 20% higher power than when in its one-core-disabled configuration.

A bug in Windows XP, which Microsoft had not fixed by press time, prevents processors from moving to their lowest idle states when USB peripherals are present, but, because I connected no USB devices, and no integrated USB peripherals, such as Webcams, the Windows power-management issue was probably not a factor in my results (Reference 4). However, although I was able to disable one of the two cores by using a BIOS setting, the second core was still present on the CPU die and, presumably, drew a tangible incremental amount of current. Therefore, a true single-core, 65-nm CPU could consume even lower power than my pseudo-single-core system did.