Evaluating value-oriented x86 CPUs: The price of falling prices

It’s unclear to what degree Intel anticipated the fiscal crisis now gripping the globe and therefore developed the economical Atom CPU as a means of preserving its market-share position. (The company publicly states that its primary intent with Atom was to broaden the application base for the x86 architecture.) It’s also unclear what negative impact on long-term corporate revenues and profits may result from Atom’s potential “cannibalization” of more lucrative Intel microprocessors. Nonetheless, any indication of robust sales is a rare and encouraging sign in today’s economic malaise. So-called netbooks currently account for the bulk of PC-market growth, and Atom is the CPU engine powering most netbook designs, so Intel deserves kudos for pragmatism in using Atom to obsolete its own CPUs instead of allowing competitors to do so.

In contrast, Via Technologies, no stranger to value-priced PCs, takes another tack. The company has long been the torchbearer of the belief that modern x86 CPUs tout capabilities that not only represent overkill for mainstream users’ computing needs, but also add costs that the manufacturers must pass on to consumers. In response, Via, through its Centaur Technology subsidiary, developed the C3 and C7 microprocessor families, which Via recently augmented with the Nano CPU. The C7-to-Nano transition markedly differs from the approach that Intel took in developing Atom, however. As I noted last spring, just as Intel was stripping out-of-order execution and other superfluous features from its previous architectures, Via’s Centaur subsidiary was poised to unleash the company’s first three-way superscalar out-of-order architecture (Reference 1).

This hands-on project is the latest attempt to answer the question: Which approach—Intel’s or Via Technologies’—is better for your design needs (see sidebar “Whither AMD?” as well as Table 1 and Reference 2). Testing stand-alone CPUs is of little practical benefit: Although the CPUs act as the brains of the systems surrounding them, the capabilities and limitations of those systems also influence them (Table 2). And raw-CPU-speed tests wouldn’t adequately represent the breadth of design requirements that contend for your attention. You must be sure to normalize performance measurements against the processor’s core clock rate. This article also provides power-consumption and price data to act as additional calibration fodder.

TEST-SUITE DETAILS
A 1.6-GHz, Atom 230-based Intel D945GCLF mini-ITX board came from Ituner Networks, whose company’s Mini-box subsidiary makes the picoPSU-90 power supply that fueled all three mini-ITX products in this project (Figure 1). Its near-twin, a 1.6-GHz, Atom 330-based D945GCLF2, came from Mobile Computing Solutions by way of eBay. The two boards look identical at first glance, aside from their CPUs, which Intel intends for use with low-cost, ac-powered PCs and other x86-based systems. A closer inspection reveals telltale differences, however: The D945GCLF2 populates, through a connector module, the s-video-output solder-pad sites on both boards, and the boards’ core-logic north-bridge-IC passive-heat-sink designs also differ.

The third mini-ITX board in this project is Via’s 1.6-GHz VB8001 Nano platform. One key difference between it and the Intel boards relates to system memory: Whereas the Intel boards incorporate only one DIMM slot supporting DDR2-533-speed maximum-memory rates, the VB8001 includes two slots that can each handle DDR2-667 SDRAMs at full speed. I wanted to test all three boards at equivalent 2-Gbyte aggregate capacities, so I populated each of the VB8001 SDRAM connectors with a 1-Gbyte DIMM. I realized that this approach might slightly increase overall system-power consumption, but the resultant dual-bank arrangement might also make overall system-memory performance better than that of the single-DIMM alternative.

In December 2007, Via unveiled the Artigo hobbyist platform, which the company based on a pico-ITX main board housing a fanless, 1-GHz C7 CPU. Artigo is diminutive: At 3.9×2.8 in., the system board takes up less than one-quarter the area of a mini-ITX equivalent and half the area of a nano-ITX board, and its surface area is only slightly larger than that of a playing card. As with the Intel boards, Artigo has a single-slot, DDR2-533-speed interface to system DRAM; however, the board supports only 1-Gbyte-maximum memory. And, as with all of the other boards, I mated Artigo with an SSD (solid-state drive) to minimize both the overall system-power consumption and the likelihood that mass-storage-read speeds would act as a bottleneck to CPU performance.

The final contestant in this benchmarking competition was a last-minute entrant. MSI based its Wind U100 netbook, like many other manufacturers’ offerings in this class, on the Atom 270, a lower-power version of the Atom 230. The Atom 270’s lower power makes it suitable for battery-fueled systems, but it is otherwise identical to the 230. This system’s processor normally runs at 1.6 GHz, but Version 1.09 of the system BIOS, which MSI released in late October, enables users to control CPU underclocking by 50% when the system is running on battery power, and it also allows for BIOS-defined overclocking by 8, 15, or 24%, when on ac power, using a simple two-keystroke combination.

As has been the case in many past projects, I used SiSoftware’s Sandra software. Conveniently for me, the company released Service Pack 1 of the 2009 Lite version just a few days before I began data collection. I first used Sandra’s various facilities during a multihour burn-in test to ensure the stability of the overclocked system. I then subjected the MSI Wind U100 to the full Sandra benchmark barrage at 1.98 GHz, or 124%, along with 800 MHz and the nominal 1.6-GHz rate. As you assess the MSI Wind-derived data, keep in mind that this system contained 1 Gbyte of SDRAM, versus 2 Gbytes with the three mini-ITX boards. However, all of the systems ran an identical Microsoft operating-system version: Windows XP Service Pack 3—the Home version on the MSI Wind and the Pro version on the others. They all also incorporated the latest BIOS and driver versions available when I did my testing in mid-November.

In reviewing the data from Sandra’s Processor Arithmetic, Processor Multimedia, and Power Management Efficiency benchmarks, the following conclusions jump out at me (Figure 2): First, as you would expect, given the similarities of the two CPUs, the data from the Atom 230 and 270 at 1.6 GHz is comparable, taking into account normal benchmark run-to-run variance. Second, the single-core Atom CPU’s performance closely scales with operating frequency; the 800-MHz-speed underclocked results are roughly half those at 1.6 GHz, with the 1.98-GHz overclocked data roughly 24% higher than the 1.6-GHz baseline. Similarly, the dual-core Atom 330 consistently delivered roughly twice the benchmark performance of its single-core counterparts at comparable operating frequencies. Also, the geriatric C7 CPU has lower performance than the 800-MHz Atom 270, even though Via’s processor operates at a 200-MHz-higher clock frequency than its Intel competitor.

The only test at which the C7 outperformed the 800-MHz Atom 270 was Multimedia Double; an analysis of the CPUs’ comparative ALU (arithmetic-logic-unit)-performance data provides a potential clue to the overall disparity, and the Atom processor’s HyperThreading support for limited parallel processing is another likely explanation. Conversely, Via’s latest Nano CPU is a robust performer, outpacing the similarly clocked single-core Atom 230 in all benchmarks despite its lack of HyperThreading-like capabilities. On the Multimedia Double test, it even significantly sped past the dual-physical-core and quad-virtual-core Atom 330.

The conclusions for the application-tailored tests, specifically those involving cryptography and both Java and .NET virtual-machine-based algorithms, largely echo those of the earlier generic-processor tests, with the exception of a few key divergences that caught my eye (Figure 3). For example, the Via Nano CPU’s earlier robust Multimedia Double capabilities did not make a repeat performance in either a Java or a .NET virtual-machine environment. Conversely, Nano performed well in the cryptography test—an expected conclusion because Via has for several product generations incorporated dedicated cryptography acceleration hardware in its CPUs. Unfortunately, a Sandra bug that I discovered—and that SiSoftware developers are now fixing—prevented me from also testing the C7 CPU with this benchmark (see sidebar “Head to the Web for the rest of the story”).

REMEMBER THE MEMORY
Given the cache-memory configuration and capacity disparities between the various CPUs in this project, as well as the system memory-configuration and capacity disparities between the systems based on them, I felt it was important to broaden my testing beyond processor-centric benchmarks to encompass memory-inclusive studies (Figure 4). Looking first at the Cache and Memory test, the overall cache/memory bandwidth of the Via Nano processor compares favorably with its Intel counterparts. However, this result is largely due to the Nano’s superior L1 cache performance. Its L2 cache capabilities, conversely, lag behind those of the clock rate-normalized Atom competitors.

Note, too, that Sandra didn’t report stand-alone data for Via’s C7 CPU’s L2 cache in either this or the Memory Latency test. “When L2 cache is equal to L1 cache, we’ve not found a reliable way to measure latency/bandwidth without the results being tainted by L1 cache,” says Adrian Silasi, SiSoftware’s lead programmer. “It is a pretty unusual configuration. Streaming to and from specific caches are really hints to the processor rather than absolute requests, and behavior varies with manufacturer and processor family.”

Focusing on the Memory Bandwidth test again reveals the Via Nano’s bandwidth inferiority. If you download the detailed background test data from the Web-site addendum to this article at the Brian’s Brain blog, you’ll see that this weakness extends across a suite of both integer and floating-point tests. Next, look at the Memory Latency tests. Focus first on the measured latencies for linear accesses. The Nano CPU’s L1 and L2 cache performance lags behind that of Atom alternatives, but its limitations are most apparent when you look at overall memory latency. When you see the background data, you’ll discover that the 256-kbyte to 64-Mbyte memory range dominates this discrepancy, again suggesting a constraint in Via’s DRAM-controller design.

This conclusion is surprising considering that the Via VB8001 board employs a dual-bank, two-slot DIMM architecture and that the Nano’s core-logic chip set runs DDR2-667 DRAM at full speed rather than at DDR2-533 speed. Speaking of system memory, the BIOS of the Intel boards allows users to optionally override the SPD (serial-presence-detection)-determined DRAM specifications. I used DIMMs capable of DDR2-667 speeds with the D945GCLF and D945GCLF2, so I manually entered the appropriate timings in the BIOS menus, but the boards subsequently refused to boot.

Note that the Via Nano’s DRAM-latency issue was linear-access-specific; the core-logic chip set conversely delivered Intel-comparable results with random accesses. The underclocked Intel Atom 270 and the Via C7 were the underachievers in this test. I suspect that this result reflects the MSI Wind’s DRAM controller’s use of consistent number-of-clock settings for relevant access parameters at various system-clock speeds, even though each clock period was twice as long in the underclocked case than with nominal clock rates.

POWER PERCEPTIONS
One commonly cited criticism of Intel’s Atom is that, whereas the CPU is thrifty from a power-consumption standpoint, its mated chip set is disproportionately current-hungry. You’ll find corroborating evidence of this opinion in the form of the lopsided passive heat sink hovering above the core logic’s north-bridge IC on one side of the CPU with the D945GCLF and D945GCLF2 boards. You’ll also find it in the disparity between CPU-only and CPU-plus-chip-set power-consumption estimates from Sandra’s various test results (Table 3). I also saw baffling evidence of it with the MSI Wind U100. I chose not to use the MSI Wind U100’s statistics that P3 International’s Kill A Watt P4400 reported because, unlike with the other boards used in this project, they included the power consumption of the optical drive and the LCD, including its LED backlight. I will mention, however, that the measured idle-current draw of the U100 when running at 1.6 GHz, 0.54A, dropped to 0.51A when the system was under the full Sandra test load.

More generally, consider that I measured the power consumption at the ac plug, not at the CPU or, for that matter, at the dc inputs to the board. As such, my tests encompass not only the power consumption of the various components on the board, but also the inefficiencies of the ac/dc and dc/dc converters, including the nonlinearity of those inefficiencies as functions of instantaneous current draw. Speaking of statistics, part of the reason that I didn’t publish Sandra’s power-normalized performance data is the inconsistent CPU-plus-chip-set power-consumption estimates various Sandra tests provided for the same CPU-plus-chip-set combination.

Even with those qualifiers, I still discerned some other interesting trends. Intel’s D945GCLF and D945GCLF2 delivered nearly identical power-consumption results in both idle and peak operating modes, an impressive achievement for the dual-core Atom 330-based D945GCLF2, considering its notably better performance than the single-core Atom 230-based D945GCLF. Second, the Via Nano CPU-based VB8001 had lower idle-power consumption but notably higher peak-power consumption than the Intel boards. Quantifying the situation, peak-power consumption for the VB8001 was almost twice the power drawn when the board was in idle mode, whereas the peak-to-idle ratios for the D945GCLF and D945GCLF2 were 1.27- and 1.29-to-1, respectively.

Keep in mind, as you compare both the idle- and the peak-power draw of the VB8001 with that of its peers, that it employed dual SDRAM DIMM banks and that those SDRAMs ran at a 25% faster data rate than did the system memory on the other boards. It probably wasn’t a coincidence that all of the single-DIMM boards drew their peak current during Sandra’s Cache and Memory test, whereas the dual-DIMM VB8001 exhibited its highest power draw during the Processor Multimedia test. Finally, although the other boards I tested eclipsed the performance of the Via Artigo kit, any system that can consume less than 9W in idle mode and 16W in peak mode—again accounting for power-supply inefficiencies—still should earn your respect.

FORECASTING FISCAL IMPACTS
Power consumption is only one of the key differentiators of raw performance; the other, price, is equally—if not more—critical. Table 1 and Table 2, respectively, provide the published Intel prices for the three Atom CPU variants that this study evaluates, as well as those for the company’s two Atom-based mini-ITX boards. Depending on your targeted application and how many CPUs you plan to buy, the price you pay might be notably less than Intel’s publicly quoted price. Remember: You also need a core-logic chip set and other system building blocks. Similarly, prices for the D945GCLF and D945GCLF2 are often significantly lower than the manufacturer’s suggested retail price. In mid-November, for example, the D945GCLF2 board was selling for $81.74, including free shipping, at Buy.com.

Alas, Via Technologies continues its longstanding and frustrating practice of declining to publish prices for its CPUs and boards. Internet searches reveal single-unit prices ranging from $270 to $330 for the Artigo kit, and company officials have suggested in interviews that the VB8001 will sell for approximately $180.

Captions
Figure 1: Intel’s D945GCLF (a) and D945GCLF2 (b) mini-ITX boards provided ideal hardware test beds for, respectively, the Atom 230 and 330 CPUs, as did Via’s VB8001 (c) for the company’s newest Nano processor. Mini-box’s compact picoPSU-90 dc/dc converter (d), in combination with a generic ac/dc 12V/5A power supply, fueled them all. Via’s Artigo hobbyist kit (e) houses a C7 CPU-based pico-ITX board (f). The MSI Wind U100’s (g) straightforward underclocking and overclocking capabilities were ideal for testing the Atom 270 CPU at various speeds, and P3 International’s P4400 Kill A Watt (h) allowed measurement of all systems’ power consumption in various operating modes.

Figure 2: Generic-processor tests in the Sandra suite produced telling statistics in Processor Arithmetic (a), Processor Multimedia (b), and Power Management Efficiency (c).

Figure 3: SiSoftware offers numerous application-specific benchmarks, including Java Arithmetic (a), Java Multimedia (b), .NET Arithmetic (c), .NET Multimedia (d), and Cryptography (e).

Whither AMD?
AMD (Advanced Micro Devices) declined my invitation to participate in this study, a decision that wasn't unexpected. Although I have a sound relationship with the company, it's not currently a major player in the netbook and other ultralow-cost-PC markets, reflecting the fact that the only credible products it can currently offer to these applications are aged, single- and dual-core K8 microarchitecture derivatives. However, at AMD's mid-November analyst meeting, the company unveiled the 45-nm-lithography-fabricated, dual-core K8-derived Conesus, which should debut in the first half of this year. If and when AMD's aspirations turn into shipped products, I'll request Conesus-based hardware and publish a follow-up report.Author Information
You can reach Senior Technical Editor Brian Dipert at 1-916-760-0159, bdipert@edn.com, and www.bdipert.com.



References
1. Dipert, Brian, “Embedded x86: keystone of your non-PC design?” EDN, May 29, 2008, pg 38, www.edn.com/article/CA6562580.

2. Dipert, Brian, “Double take: Reassessing x86 CPUs in embedded-system applications,” EDN, April 27, 2006, pg 69, www.edn.com/article/CA6325588.

Click here for the illustrations:

Figure 1, Figure 2, Figure 3, Figure 4, Table 1, Table 2, Table 3