x86 processors: Continued innovation is a welcome contradiction

In mid-2008, when EDN last looked at the broad x86-processor product segment in detail, the world economy was six months into a deep recession (Reference 1). Roughly six months later, worldwide economic conditions remained grim (Reference 2). Although government officials now claim that the recession is over, key economic indicators, such as the unemployment rate, should deliver underwhelming statistics for some time to come.

You might expect, therefore, that the semiconductor industry would mirror the broader economy’s malaise; this scenario has indeed occurred in many technology and product sectors. The x86 CPU business has bucked the general trend, however, as continued R&D investment and resultant new-IC output demonstrate. “Irrational exuberance,” a term that former Federal Reserve Board Chairman Alan Greenspan coined in 1996, has thankfully not been evident. However, the microprocessor and core-logic-chip-set suppliers know full well that, with the rapid evolutionary pace of the PC industry, to tangibly slow down in the short term might be sufficient to ensure demise in the long term. This insight is evident in the continuation of their prerecession momentum.

Desktop expansion

Recent analyst reports estimate that Intel holds more than 80% of the overall x86 CPU market and an even higher percentage in key product segments, such as mobile systems and servers. The company also has a higher associated R&D budget and more employees than its key competitors. It is perhaps unsurprising then that most of the new- product roll-outs over the past several years have come from Intel. In 2008, the company launched the Core i7 CPU, its initial product employing the Nehalem microarchitecture. Intel has since substantially broadened the Nehalem line, and you must differentiate among silicon platforms and resultant products to comprehend the full scope of the roll-out.

First, recall Intel’s “ticktock” product-and-process cadence, a term that company President and Chief Executive Officer Paul Otellini used in 2006. Intel aspires to reduce overall risk by making major product and process transitions at different times. In the “tick” cycle, the company shrinks a product’s lithography to the next process node, making only minor feature alterations in the process. The subsequent “tock” step maintains the same process node but implements more substantial product changes.

The late-2008 introduction of the Core i7 CPU, a high-end manifestation of the Bloomfield silicon platform, reflected a tock on Intel’s 45-nm process (Figure 1). Its innovations versus earlier Core-microarchitecture-based products include three channels’ worth of integrated memory controllers; the QuickPath Interconnect, an AMD (Advanced Micro Devices) HyperTransport-reminiscent core-to-core and CPU-to-CPU mesh; the resurrection of hyperthreading core virtualization; a trilevel integrated-cache architecture; and Turbo Mode, which automatically boosts the clock speed of active CPU cores in situations in which some cores are not in use, maintaining the chip’s thermal envelope and boosting overall performance.

Late last year, Intel launched Lynnfield, its second 45-nm, Nehalem-microarchitecture-based silicon tock offering, this time for mainstream-desktop-CPU platforms. Lynnfield discards one of Bloomfield’s DRAM channels, instead integrating 16 lanes’ worth of PCIe (Peripheral Component Interconnect Express) 2.0 I/O that previously resided in a separate north-bridge core-logic IC. The interface to the remainder of the external core logic is also the 2-Gbyte/sec Direct Media Interface bus; Bloomfield in contrast uses the more-than-10-times-faster QuickPath bus. Lynnfield translated into both Core i7 and Core i5 product proliferations; Core i5 disabled hyperthreading support.

Intel recently executed its next tick by migrating to a 32-nm process commensurate with the unveiling of the low-end Clarkdale CPU. Clarkdale comes in both Core i5 and Core i3 product flavors; some variants disable not only hyperthreading but also turbo-mode support. Instead of relying exclusively on stand-alone or core-logic-embedded external graphics-accelerator circuitry, Clarkdale places graphics-processing capabilities alongside the CPU in a multidie, single-package arrangement. Clarkdale also migrates the system-memory controller off the CPU and onto the companion die, potentially degrading performance at the trade-off of lower CPU silicon cost. The technology treadmill rolls on; Intel demonstrated its next tock, the 32-nm Sandy Bridge microarchitecture, in functional-silicon form at last fall’s Developer Forum. The company showed off a 22-nm test wafer at the same venue.

Intel’s primary x86 CPU competitor, AMD, has also over the past several years been fairly busy in this product segment, befitting the company’s comparatively strong market-share position in desktop PCs versus other markets. After a problem-plagued roll-out of the 65-nm Phenom CPU in late 2007, complete with poor yields and L3-cache-controller bugs, the subsequent 45-nm Phenom II transition beginning in January 2009 has seemingly gone more smoothly. Phenom II boosts the L3-cache allotment over that of its predecessor and tweaks other minor architectural details. As with Phenom, AMD sells Phenom II not only in full quad-core form but also in triple-core X3 and dual-core X2 variants. These product spins also contain four cores on each die, but AMD disables one or two of them, presumably because they don’t pass functional or speed testing. The company is attempting to maximize usable product yield from each manufactured silicon wafer, but this strategy is costly and less than ideal.

Therefore, for mainstream-desktop-PC applications, AMD last summer rolled out the 45-nm Athlon II product line. Whereas the company derived the original Athlon families from the K8 and earlier microarchitectures, Athlon II evolves from the same K10 core foundation as Phenom and Phenom II. However, as a die-slimming move, AMD offers Athlon II only in dual-core form; the company also deleted Phenom II’s L3-cache array, compensating somewhat by boosting the L2-cache allotment. AMD also still sells single- and dual-core, 65-nm Athlon products. More generally, the company has attempted to counterbalance a comparative dearth of fundamental product proliferations through a blizzard of clock-speed and operating-voltage spins, striving to satisfy a plethora of product-performance-versus-price expectations with comparatively few silicon foundations.

Mobile migration

As anyone who has recently visited computer retailers or scanned their advertisements realizes, desktop-computer sales are at best stagnant, whereas mobile-computer sales are exploding. Intel reportedly controls approximately 90% of this market, so it’s no surprise that the company has focused significant attention in this area. Most Intel-based notebook computers still use so-called CULV (consumer-ultralow-voltage) CPUs employing the previous generation’s Core microarchitecture, although the company last fall began shipping mobile Core i7 CPUs it derived from Clarksfield, a notebook-intended variant of the Nehalem-based Lynnfield. More recently, Intel unveiled Arrandale, a mobile version of Clarkdale. The current modern differentiation between desktop and mobile CPUs is nebulous at best. It reflects operating-voltage and clock-frequency variations but little if any underlying circuitry differences because power consumption is a critical consideration in all market segments (see sidebar “Server spins”).

To date, AMD has not attempted to aggressively shoehorn Phenom, Phenom II, or Athlon II CPUs into notebook computers, relying instead on power-consumption-optimized earlier-generation Athlon products. The company prefers to promote and sell them in processor-plus-chip-set form. Take, for example, the Neo platform, which AMD unveiled at the January 2009 Consumer Electronics Show along with system partner Hewlett-Packard (Figure 2). AMD, by virtue of its 2006 acquisition of ATI Technologies, now touts formidable stand-alone and core-logic-integrated graphics capabilities. And AMD isn’t shy about pointing out its claimed advantages over Intel’s alternatives as they relate to such topics as Windows Vista and Windows 7 UI (user-interface) enhancements, mainstream 3-D gaming scenarios, and increasingly important hardware-assisted decoding of leading-edge video codecs.

Via Technologies, meanwhile, has unveiled three variants of its latest-generation, 64-bit, single-core Nano CPU. The 1000 and 2000 series of products are identical from the silicon and microcode standpoints; the only differentiation is with respect to the marketing moniker. The newer 3000 series, on the other hand, represents a more involved circuit redesign, adding support for SSE4 (streaming single-instruction/multiple-data extension version 4) and with claimed higher performance and lower power consumption. The long-promised dual-core variant of the Nano family remains absent; the latest company road maps forecast its arrival to occur no earlier than mid-2010. Nonetheless, Via has secured several design wins with Tier 1 notebook-PC OEMs, such as Lenovo and Samsung. Via also continues to sell its earlier-generation C7 processors.

Fiscal evolution

The term “netbook” represents Intel’s attempt to separate the systems from notebooks (see sidebar “Handsets and other CE hopes”). Netbooks contain low-cost Atom microprocessors, whereas notebooks include higher-end, higher-profit, and higher-revenue-garnering Intel CPU alternatives. A third category, the smartbook, includes 3G (third-generation) cellular-data subsystems and often offers carrier subsidization. Processor differences aside, Intel and partner Microsoft have undertaken other netbook-versus-notebook differentiation steps. These steps include restricting to 2 Gbytes the maximum system memory that the associate core logic supports, encouraging their customers to further decrease the installed system memory to 1 Gbyte, and constraining the screen sizes and resolutions that netbooks can implement. Historically, Microsoft’s motivation has been to minimize the number of less lucrative Windows XP licenses that it sold into netbooks versus those for the more fiscally attractive Windows Vista in notebooks. Nowadays, its intention is to steer as many licensees as possible away from the netbook-tailored Starter Edition of Windows 7 and toward more expensive variants of the OS.

Intel in December formally unveiled the latest Atom CPU, Pineview—or Pine Trail in CPU-plus-chip-set-platform lingo—which exemplifies this segmentation trend (Figure 3). Pineview integrates the graphics core and memory controller that the earlier-generation Diamondville device implemented on a separate north-bridge core-logic IC. Pineview’s approach has battery-life advantages, especially when you consider that the earlier 90-nm-process-based 945GC and 130-nm ICH7 consumed far more power than did the 45-nm-derived Atom N- and Z-series CPUs.

However, Intel also implemented the Pineview Atom CPU on a 45-nm process, which translates to an increase in transistor count over precursor Atom CPU generations. To minimize die size, Pineview’s graphics core has fewer features than that of the 945GC. This seeming step backward may make you wonder why else Intel might have made this integration move. Performance is a likely reason; although the CPU core itself remains the same, the tight coupling to the graphics and memory subsystems is preferable to the earlier device’s separation through the front-side bus. Also recall that the graphics core employs system memory as its frame buffer. This time around, the device supports 667- and 800-MHz DDR2 SDRAM versus only 533-MHz memory on the 945GC. Unfortunately, Pineview’s video outputs have maximum resolutions of only 1366×768 pixels from digital ports and a maximum of either 2048×1536 or 1400×1050 pixels with an analog VGA connection.

Chip-set competition

Pineview’s tight coupling between the CPU and the GPU (graphics-processing unit) also conveniently—at least for Intel—shuts out third-party graphics products, which previously connected to the Atom processor through that same front-side bus. Instead, third-party graphics chips must tether to Pineview through the NM10 companion chip’s PCIe lanes, several bus “hops” away from the processor. System memory, which such a GPU might prefer to also employ as its frame buffer, is also several hops away, resulting in either poor performance or the necessity for a costly dedicated frame buffer that directly connects to the third-party GPU.

Pineview’s integration, reminiscent of that in the Clarkdale and Arrandale CPUs, brings to the forefront a long-standing feud between Intel and Nvidia (Reference 3). Nvidia’s Ion chip, an alternative to Intel’s two-chip 945GC-plus-ICH7 approach, muddles the distinction that Intel strives to preserve between netbooks and notebooks. Ion uses the same IC that Apple’s latest Intel-based MacBook laptops and entry-level Mac mini desktop computers use. It has secured a few key design wins in portable systems alongside Intel’s first-generation Atom CPUs, as well as in a host of so-called nettops for home-theater-PC and other niche applications (Figure 4).

Ion’s fundamental appeal, aside from its single-chip function integration, is that it has notably better graphics- and video-processing capabilities than Intel’s alternative. Ion offers some degree of hardware-accelerated decoding for advanced video-compression algorithms, such as H.264, which Adobe Flash is adopting, and VC-1, which Netflix’s Watch Instantly online service employs. Blu-ray discs also use both H.264 and VC-1; Ion handles most of the processing itself instead of burdening the CPU with the task. Nvidia has long publicly complained, however, that the Atom pricing strategy is competitively unfair in that Intel charges significantly more for the stand-alone CPU than when it sells it with the 945GC and ICH7. Indicating the potential reality behind Nvidia’s claim, systems based on Ion tend to be significantly more expensive than their “generic” Atom counterparts. It’s unclear, however, whether this differentiation reflects bill-of-materials costs or PC OEMs’ desires to use potential customers’ perception of the superiority of an Ion-inclusive system to boost its price.

Speaking of pricing, even if Nvidia figures out how to make a cost-effective offering with compelling performance with its upcoming Ion 2 chip for Pineview, the product’s success isn’t assured. Pineview alone has sufficient graphics horsepower to run Windows 7’s rich GUI (graphical user interface), for example. You may thus conclude that video is the primary motivation to append Ion 2 to Pineview. An even more cost-effective silicon alternative exists, however, in Broadcom’s Crystal HD (high-definition) video-decoder IC. The company initially unveiled the device several years ago, but it has only now hit its stride in conjunction with the ascendancy of the Atom CPU. Intel prominently showcases Crystal HD in its Pineview promotional materials. And OEMs sell Crystal HD-augmented systems for only $30 or so more than their Atom-only counterparts, suggesting that Broadcom is aggressively pricing the part to carve out a market and fend off the Nvidia alternative. To buffer itself against Atom-only business dependence, Nvidia also plans to develop Ion variants for use with Via’s Nano CPU series.

Consumer protection

Ion is only one of several areas of disagreement between Intel and Nvidia. Nvidia also has yet to obtain a license for either the QuickPath Interconnect or Direct Media Interface buses, meaning that the company cannot legally sell chip sets for Nehalem-generation Intel microprocessors. More generally, Nvidia is struggling from a business standpoint, with its DirectX 11-supportive graphics chips notably delayed and with AMD’s microprocessor customers also tending to select AMD/ATI chip sets and discrete GPUs.

The disputes between the two companies are among several motivations that the US FTC (Federal Trade Commission) mentioned when in mid-December it filed a lawsuit against Intel for alleged anticompetitive practices. This latest legal setback is only one in a series that Intel has suffered in the past year. The European Union also last May fined Intel $1.45 billion, which the company agreed to pay, although it vowed to appeal the ruling. In November, New York State filed an antitrust lawsuit against Intel, and Intel in the same month paid AMD $1.25 billion to settle long-standing litigation between the two companies.

Nvidia may be closely following the evolution of the FTC action for more reasons than concern about chip sets and graphics. An unconfirmed long-standing rumor suggests that the company has for some time been internally developing an x86-microprocessor architecture. Both design-team acquisitions and official—albeit obscure—company comments bolster this rumor. However, Nvidia has not yet obtained permission to sell chips employing the architecture. If the x86-by-Nvidia rumor is true and if the FTC’s legal aspirations are successful, one possible remedy would be for the FTC to require that Intel grant Nvidia an x86 license.