Medfield: Will Intel's Long-Standing x86-in-Smartphone Aspirations Finally be Fulfilled?

Intel has been striving to shoehorn the x86 CPU architecture into handheld communications and computing devices ever since the company began publicly discussing the Atom architecture in late 2007. First- and second-generation Atom-based CPUs and associated core logic chipsets found predominant success in netbooks. But Intel also targeted low-voltage and reduced-clock-speed variants (in some cases also swapping out internally developed graphics accelerators for PowerVR cores licensed from Imagination Technologies) at tablets, MIDs (mobile Internet devices, aka UMPCs, ultra-mobile personal computers) and other pocket-able and otherwise portable widgets, with limited success.

Nonetheless, with Moorestown, introduced in May 2010, Intel began publicly talking up its cellphone aspirations. A lithography shrink to 45 nm (for the primary SoC; the companion chip was still fabricated in 65 nm at TSMC) dropped the chipset count to two ICs, not counting the cellular baseband and other analog-heavy peripherals. And, by dropping PCI bus support and other legacy standards, required for full Windows compatibility but unnecessary for mobile operating systems, Intel signaled it was getting serious and dedicated development resources to a phone-tailored product line.

Unfortunately, Moorestown was still too power-hungry, compared to competitive offerings from ARM licensees such as Nvidia, Qualcomm and Texas Instruments, to garner much interest from handset manufacturers. Similarly, the comparatively trailing-edge process lithography on which Intel fabricated it excessively bloated the total chip count required to implement a system design. LG Electronics worked with Intel to develop a handset prototype, but LG's subsequent corporate fiscal contractions killed this particular project before it came to production. And Nokia's subsequent decision to transition to (and focus on) Microsoft's Windows Phone O/S, to the demise of the MeeGo (which had merged Intel's Moblin and Nokia's Maemo mobile O/Ss), doomed Intel's x86 aspirations with this particular phone manufacturer as well.

As Intel's shown repeatedly in the past, however, it has the deep pockets and willingness to invest for the long term in areas that it views as strategic. To wit, the company created one of the biggest news "splashes" of January's Consumer Electronics Show when it formally introduced the Atom Z2460 SoC, previously known by the code name Medfield (Figure 1):

Figure 1: A 32 nm process foundation leads to Medfield's increased integration.

By transitioning the Atom microarchitecture to 32 nm process, Intel was able to combine the prior-generation multi-chip suite into a single-die offering. And further upping the integration ante, Intel sells the Atom Z2460 in a stacked-die PoP (package on package) configuration that also incorporates the system DRAM. The "Saltwell" Atom core is essentially unchanged from its late-2007 "Bonnell" origins, aside from the die size, performance, power consumption and yield improvements deriving from a multi-generation lithography shrink. Note that although the Atom Z2460 is listed as running at up to 1.6 GHz, this particular clock speed reflects the chip's "Turbo" mode, which isn't sustainable for long periods of time due to peak TDP (thermal design power) constraints. The sustained peak clock speed is 1.3 GHz, and the minimum operating clock rate is 100 MHz; the core automatically ramps up and down between those extremes as software demands dictate and in 100 Mhz "step" increments.

Supplementing the base CPU capabilities are numerous on-die coprocessors, beginning with the PowerVR SGX 540 GPU running at up to 400 MHz, which you can also harness to tackle other parallel computing tasks beyond graphics through an API such as OpenCL. Intel also includes dedicated video encode and decode function blocks, licensed from Imagination Technologies, along with an image processor deriving from the company's year-ago acquisition of Silicon Hive. Note, too, Medfield's 512 KByte L2 cache supplemented by dual 32-bit low-power DDR2 SDRAM interfaces. Nvidia's Tegra 2 SoC, in contrast, bundles 1 MByte of L2 cache, but couples it to only a single-channel 32-bit LPDDR2 SDRAM interface, a design decision that the company surprisingly carried forward to its next-generation Tegra 3 design.

Nvidia’s decision to integrate twice the L2 cache in the Tegra 2 compared to the Intel Z2460 is understandable when you consider that Nvidia's device is a dual-physical-core configuration.  And that fact begs the question of how Intel thinks it'll be able to compete, with a single-core Atom offering, against the now-common two-core SoCs of ARM's licensees. Part of the answer involves comparative clock frequencies; it's still rare to find a mainstream ARM Cortex-A9-based device that runs at greater-than-1 GHz speeds. And although synthetic benchmarks are crafted to leverage as many cores as are available in a given handset, it's also rare to find a realistic mobile O/S usage scenario that fully maxes out a single core, far from multitasking (or multi-threading) to simultaneously harness multiple cores.

Comparatively smaller system memory allocations, along with far more stringent power consumption limitations, are the primary reasons for this CPU resource-usage disparity between handsets (and tablets) and conventional computers. And note, too, that Intel's Z2460 is a dual-core CPU, at least of a virtual-core sort, when conditions warrant and are amenable to it. The company's HyperThreading simultaneous multithreading architecture, which first appeared on the Pentium 4 processor and has been enhanced through several generations of desktop, laptop and now handheld CPUs, presents itself as two distinct cores to an operating system and application software and can act as such as long as contention for main execution resources does not exist.

The result, per data that Intel published at CES in January, is that the company believes the Atom Z2460 is competitive both in performance and power consumption with the ARM-based SoCs found in today's handsets (Figure 2):

Figure 2: Intel believes that Atom Z2460-based designs can hold their own against today's shipping smartphones, in terms of both performance and power consumption.

Respected technology analysis site AnandTech had the opportunity to run some preliminary tests on an Intel Medfield-based reference design (Figure 3) at CES, compared them against benchmark results it had previously logged on other hardware, and came away favorably (albeit cautiously) impressed.

Figure 3: Intel's Medfield reference design comes in sleek, production-ready packaging that several announced customers have heavily leveraged.

Keep in mind as you look at AnandTech's numbers (Figure 4) that the Intel reference design was running Android v2.3 ("Gingerbread"), while the competitor handsets and tablets were largely powered by more modern and performance-optimized Android v3.x ("Honeycomb") and Android v4.x ("Ice Cream Sandwich") O/Ss, along with Apple's iOS.

Figure 4: Preliminary hands-on testing results bear out Intel's rosy Medfield predictions.

Speaking of Android, its virtualization-or-not attributes will be a key factor in how speedy-or-not an Intel Atom-based handset will seem to be in comparison to an ARM-based counterpart. Since de-emphasizing MeeGo, Intel has wisely redirected its development energy toward Google's O/S. At Intel's September 2011 Developer Forum, for example, Google's Andy Rubin emphatically stated that the Android 4.0 (and beyond) SDKs would include full Intel CPU architecture support. This is especially good news to Intel for new-application development.

But what of the all-important existing Android apps? If they've relied exclusively on the SDK, they should present no problems for an Intel-based system, since the code is compiled on the fly and run through the hardware-isolating Dalvik virtual machine common to both x86 and ARM CPU implementations. If, on the other hand, they leveraged the Android NDK for performance optimization reasons, the ARM-native code generated by it would seemingly be x86-incompatible. However, Intel has developed a dynamic binary translation software layer that will bridge the short-term code divide, albeit with both performance and power consumption efficiency penalties, until the developer is able to recompile the application in a way that directly supports x86.

Also at CES, Intel unveiled partnerships with Motorola, which plans to ship Medfield-based phones beginning this summer followed shortly by tablets, and Lenovo, which will initially target the China market with its Medfield-based cellular handset offerings. And at Mobile World Congress a few weeks ago, Intel further extended its winning streak with several other notable partnerships. For example, European cellular service provider Orange will sell a branded handset based on Intel's reference design (Figure 5), as will Indian handset manufacturer Lava. And Intel also unveiled a multi-year partnership with Chinese hardware developer ZTE, similar to the one it had previously announced with Motorola.

Figure 5: "Intel Inside" comes to the European smartphone market with Orange's planned offering.

At Mobile World Congress, Intel also announced two more Medfield-based chips, both due to enter production next year (Table 1). The Atom Z2000 is a down-binned variant of the Atom Z2460, running at 1 GHz CPU and 320 MHz GPU peak clock speeds, and with HyperThreading support disabled. Conversely, the high-end Atom Z2580 doubles up both the CPU physical core count to two physical (four virtual, via HyperThreading) and the GPU core count to two (via a migration to the next-generation PowerVR SGX 544MP2). Clock speeds also increase, to 1.8 GHz max for the CPU and 533 MHz for the GPU.

Table 1: Feature set and performance comparisons of Intel's first three Medfield-based SoCs.

On the one hand, it might be easy to dismiss Intel's latest stab at the smartphone market as "too little, too late." After all, the new Z2460 is arguably competitive with ARM-based counterparts that have been available for some time, and next-generation ARM-based products are looming on the horizon. Nvidia's quad-core Cortex-A9 Tegra 3 is already in production, with Qualcomm now sampling Krait-based SoCs and Texas Instruments sampling its dual-core Cortex-A15 initial offerings. However, while the high-end system designs based on these leading-edge SoCs may deliver bragging rights, they don't deliver much volume in comparison to the mainstream smartphones that Intel's going after, particularly in high-growth markets such as China and India.

Look beyond the initial three Medfield products and Intel's competitive prospects are even more promising. The company's 22 nm process, which will ramp into initial volume production next quarter in the form of Intel's Ivy Bridge desktop and laptop CPUs, is equally amenable to Atom-based SoCs, and Intel has already publicly promised to rapidly migrate Medfield descendants there. If and when it does, they'll represent a one- or more-generation lithography advancement beyond what ARM foundries are able to deliver in that timeframe, with commensurate benefits in cost, transistor budget, power consumption and performance.

Consider, too, Intel's acquisition of Infineon's wireless solutions business, announced in August 2010 and completed in January 2011. It gave Intel access to Infineon's cellular baseband technology, which Intel plans to integrate within near-future Atom-based SoCs. This accomplishment will match the integration capability that Qualcomm currently possesses and that Nvidia also aspires to deliver in the future via its mid-2011 acquisition of Icera. And consider, too, that Intel likely has a next-generation Atom microarchitecture well along in development at this point, intended to keep pace with the Cortex-A15 and Krait alternatives.

In traditional PC businesses, Intel is the 10,000 pound gorilla, and its customers use AMD (and are considering using ARM) CPUs in part to counterbalance Intel's dominance and "keep it honest." Ironically, in contrast, in smartphone and tablet markets Intel is the "new kid in town." Mobile device manufacturers yearning for an alternative to ARM's licensees may not have been able to credibly consider x86 before, due to performance, power consumption and other shortcomings. But going forward, they may be increasingly tempted to do so.