For much of the past year we've been hearing that Intel's Atom for Windows 8 tablets is going to be the one to beat. Ivy Bridge (and later, Haswell) will exist at the high-end of the Windows 8 tablet space but if you want Intel's answer to ARM based Windows RT tablets it's going to come from the Atom lineup. It's still too early to talk about pricing, but expect Atom based Windows 8 tablets to exist in the sub-$600 space. Exactly how low they go will depend entirely on what the OEMs decide to ship at. My hope is for prices to start around $399 rather than much higher but we'll see just how seriously the Windows 8 OEMs are going to take this launch.

Intel's Atom for Windows 8 tablets has gone under the codename Clovertrail for quite some time. Today it gets official: the first Clovertrail SoC is Intel's Atom Z2760. Following similar naming to Medfield's Atom Z2460, there's a lot more that's shared between Clovertrail and its smartphone counterpart.

Similar to its approach with Medfield, Intel built a Windows 8/Clovertrail form factor reference design. The 8mm thick 10.1-inch tablet was built by a single ODM, but four manufacturers total are allowed to reuse those designs as they see fit. The FFRD approach isn't as necessary here as it was with Medfield because most of the traditional PC OEMs are already used to working with Intel. Given Intel's intense focus on driving platform power down however, building a reference design that others can follow makes a lot of sense. We'll actually see even more of this with Haswell next year.

From an availability standpoint we'll see the first Atom based Windows 8 tablets this fall. Intel announced design wins with Acer, ASUS, Dell, Fujitsu, HP, Lenovo, LG, Samsung and ZTE. The thinnest tablets will measure 8.5mm in thickness and weigh as little as 1.5 lbs. Not all Atom based Windows 8 tablets will be available on October 26th, but we'll see them starting a few weeks later.

Atom's big promise of course is ARM based tablet pricing with full x86 backwards compatibility, allowing the use of virtually all of your existing legacy Windows 7. It's a pretty compelling sell. If Intel can deliver the same (or better) performance/battery life compared to the ARM based Windows RT tablets, while preserving backwards compatibility Clovertrail would seem like a no-brainer. Simple solutions are rarely so simple, so we'll have to see this one play out in the market to be sure. Until then, we have some much needed architectural detail on Clovertrail and the Atom Z2760.

The Architecture

Medfield vs. Clovertrail
  Intel Atom Z2460 Intel Atom Z2760
Platform Codename Medfield Clovertrail
OS/Platform Target Android Smartphones Windows 8 Tablets
Manufacturing Process 32nm SoC (P1269) 32nm SoC (P1269)
CPU Cores/Threads 1 / 2 2 / 4
CPU Clock up to 2.0GHz up to 1.8GHz
GPU PowerVR SGX 540 PowerVR SGX 545
GPU Clock 400MHz 533MHz
Memory Interface 2 x 32-bit LPDDR2 2 x 32-bit LPDDR2

Clovertrail is the platform name, the single-chip SoC is called Cloverview. The architecture is very similar to Medfield's Penwell SoC. The most obvious differences are in the CPU and GPU configuration. Cloverview features two 32nm Saltwell Atom cores instead of one in Medfield/Penwell. These two cores run at up to 1.8GHz, a slightly lower clock speed compared to the Penwell that ships in Motorola's RAZR i. Each core has its own private 512KB L2 cache.

The GPU is a PowerVR SGX 545 from Imagination Technologies, clocked at a very high 533MHz. Direct3D feature level 9_3 is officially supported. Intel claims that the GPU is fed by a better engine in Clovertrail/Cloverview than it is in Medfield, however any additional details were impossible to come by.

The memory controller remains a dual-channel LPDDR2-800 design. Most tablets will see two 1GB devices populating the channels. Package on package (PoP) stacks will be used for DRAM and SoC integration, similar to what you see in a smartphone.

The other changes are more subtle. Video encode/decode is handled by IP licensed from Imagination Technologies, however the encoder is newer than what was used in Medfield. Clovertrail uses the same Silicon Hive ISP from Medfield. It supports two cameras (2MP/8MP) and burst mode.

The platform supports true connected standby, meaning Intel's new S0ix sleep states (similar to what was announced in Haswell). I realized I haven't yet detailed what these mean yet but in short on DC power you can expect polling roughly every 30 seconds for new data (incoming emails, tweets, etc...) delivering an experience somewhat similar to a smartphone. Off-SoC device drivers need to support Windows 8 run time power management (RTPM) to support these new low power sleep states. Intel claims that in its lowest platform sleep state (S0i3) the SoC's power consumption is below 2mW.

Connected standby is only currently supported by 32-bit Windows 8, so although Clovertrail supports x86-64 the platform will launch as 32-bit only. There's no support for alternate OSes at this point.

The SoC doesn't support SATA, just eMMC like most other smartphone/tablet SoCs. This is a bit of a disappointment as most eMMC controllers are pretty bad, but Intel tells us they've been working to improve things with the controllers that are out there.

There's no USB 3.0 support, Clovertrail just supports two USB 2.0 ports (OTG + xHCI, although OTG isn't supported by Windows 8). OEMs can obviously integrate hubs in any docking stations they may build.

On Pricing and ARM

Intel claims that there's no reason that Atom based Windows 8 tablets, from a hardware bill of materials perspective, should be any more expensive than their ARM based counterparts. The important takeaway is that Intel is significantly reducing the price of the Atom Z2760 due to competitive pressure from ARM. Most ARM smartphone SoCs seem to be priced in the $15 - $30 range, and I'd expect the Z2760 to fall somewhere in that range. Intel has shipped cheap CPUs in the past, but I don't know that they've ever shipped something this cheap. ARM's impact on Intel is measurable, it is the new AMD.

On Performance and Power

Microsoft isn't allowing any hands on performance of Windows 8/RT tablets yet so we don't have any of our own performance data to share. Intel did share some SPEC CPU2000 data it ran on its own with competing platforms. The data below wasn't generated by us so take it with a gigantic grain of salt:

Intel Generated SPEC CPU2000 Comparison - Normalized to Snapdragon S4
  SPECint SPECint_rate
Qualcomm Snapdragon S4 (1.5GHz) 1.00 1.00
Dual-Core 40nm ARM Cortex A9 (1.8GHz) 1.14 1.14
NVIDIA Tegra 3 (1.3GHz) 0.86 1.25
Intel Atom Z2760 (1.8GHz) 1.20 1.54

Everything here is normalized to the performance of Qualcomm's Snapdragon S4 (dual-core Krait, 1.5GHz). Basically it shows a tangible advantage for Clovertrail. That's not too hard to believe given what we've seen in the phone space, although we'll have to wait and see once we get our hands on final hardware. Sunspider scores at or below 800ms should be possible as that's what we've seen on preproduction hardware already.

We've already established that Medfield is competitive from a power standpoint with ARM based SoCs. It doesn't offer the best power characteristics, but it's hardly the worst. Middle of the road is the best way to put it. We don't have (and aren't allowed to have) a Clovertrail based Windows 8 tablet yet, so we'll have to reserve full analysis until then. However, Intel did share some early data with us from its own testing that points to Clovertrail battery life being competitive with other platforms:

Intel Generated Power Consumption Comparison
  Idle (Screen On) Web Browsing HD Video Playback
ASUS Transformer Pad Infinity 2.7W 3.4W 3.1W
Apple iPad 2 2.5W 2.6W 2.5W
Apple iPad (2012) 4.3W 4.5W 5.9W
Intel Clovertrail FFRD 2.3W 2.8W 3.0W

Intel standardized on 200 nits for all of its battery life tests, however I wasn't allowed to study/mirror the workloads and test procedure. The data looks good for Intel. Clovertrail's power consumption appears to be lower than NVIDIA's Tegra 3 and a little worse than Apple's iPad 2. This all looks quite plausible, I'm curious to see how power consumption would compare in Intel's tests to Qualcomm's S4. We'll find out for ourselves in due time.

Final Words

Intel was pretty light on Clovertrail details other than what we've published here. The real work begins once we start getting hardware late next month. The biggest question is really whether or not the OEMs will get pricing right for these tablets. An affordably priced Windows 8 tablet running Clovertrail can be very compelling for someone looking to carry a single device instead of a tablet + notebook. As with most things however, I am worried that we'll have to wait at least one more generation for perfection. I can't tell if I'm being cynical or realistic. Let's hope I'm surprised come October 26th.

POST A COMMENT

78 Comments

View All Comments

  • somata - Friday, September 28, 2012 - link

    SPECint typically does a decent job of gauging performance on general-purpose code, and the numbers presented seem believable to me. I too was surprised that Krait didn't perform better, but Geekbench also shows Krait performing worse than expected (vs the A9), even if you just focus on the single-threaded tests. Both SPEC and GB are suites and not just single benchmarks, so this makes me think perhaps Krait isn't the A15 competitor it's made out to be. Nevertheless, at least in GB, Krait tends to just edge out Atom's per-clock performance in the single-threaded tests, on average.

    Apple's A6 cores OTOH are surprisingly powerful. In fact, I predict they'll turn out to be more than a match for Cortex A15, both in IPC and perhaps even in absolute performance once they get clocked higher in the next iPad. Likewise, given that Intel's in-order Atom is still competitive with the latest out-of-order ARM cores, and how easily Intel's current out-of-order designs annihilate even the A6, I think Intel will have little trouble overtaking ARM from a performance standpoint once the next-gen Atom is released.
    Reply
  • Wilco1 - Sunday, September 30, 2012 - link

    Spec is well known as the most gamed benchmark as it depends on the compiler and options used. If you think Intel did this test fairly then think again... If you look at less biased benchmarks then you'll see a Cortex-A9 beating Atom by a big margin. For example in Geekbench a 1.4GHz Galaxy S3 easily beats a 1.6GHz Medfield on either single or multithreaded benchmarks.

    Eg. http://browser.primatelabs.com/geekbench2/1090947 (Z2460 score 882)
    and http://browser.primatelabs.com/geekbench2/1107936 (Exynos4 score 1773)

    So your conclusion that Intel's in-order core is competitive with ARM out-of-order cores is based on Intel's marketing claims rather than reality.

    Krait is a little faster than an A9 clock for clock. However given that it turns out the A6 actually runs at 1.2-1.3GHz, it will come close but I don't expect that it will match the A15 (which is about 50% faster than an A9 clock for clock).
    Reply
  • somata - Monday, October 01, 2012 - link

    Fair point about SPEC, but like I said, in my experience Geekbench results lead to a similar conclusion. The results you provide paint Atom in an especially bad light since you're comparing a 1C Atom to a 4C Cortex A9. That may be a fair comparison as far as overall platforms go (for now), but I'm only referring to the strength of the core; going multi-core is trivial.

    I never said that Atom necessarily offers better performance than competing ARM cores, only that it's competitive, which it certainly is. Notice that Krait somehow loses most tests to slower-clocked A9:
    http://browser.primatelabs.com/geekbench2/compare/...
    Against Atom it's basically a wash (again focusing on single-threaded performance). Also, Android on x86 is AFAIK still pretty immature, so I'm not sure how much the results you cite can be trusted. In Linux the same Atom would score a bit over 1000 overall and win several tests against the A9.
    Reply
  • Wilco1 - Monday, October 01, 2012 - link

    Comparing single-core Z2460 with a quad-core Exynos is reasonable as those CPUs are used in the latest phones. The Exynos is 33% faster on the integer benchmarks and 68% on the floating point (single-threaded only, at the same frequency). I do not call this competitive. Dual cores and a higher turbo won't be enough to make it competitive against current Exynos4/Tegra3, let alone the A15-based next generation.

    Yes, it seems current Krait's are hobbled by a slow memory system or a bad memory controller, the memory scores look really bad. QC have to fix this and increase frequency to ~2GHz if they want to stay competitive.

    Android is based on Linux, so I'm not sure why you think Linux would improve scores. It's true that netbook Atoms have better scores, but these have a much faster memory system and use far more power. I've seen Phoronix results comparing OMAP and Exynos development boards with netbooks on Linux, and the results are the same: an in-order Atom is only competitive if it is clocked much higher. For mobiles and tablets this isn't feasible because of power.
    Reply
  • somata - Tuesday, October 02, 2012 - link

    Indeed I was using an Atom N455 as a stand-in for the Z2460 since I believe both share nearly the exact same microarchitecture. I thought maybe I overlooked differences in the memory interface, but both have 6.4GB/s of available bandwidth. With that in mind, I can't see why the N455 should have ~20% better IPC than the Z2460 in Geekbench other than the OS difference...

    Then again, I'll grant you there may be lower-level differences in the memory/cache interface that negatively affect performance. If I concede that the Z2460 is accurately portrayed in the handful of available GB results, then its per-clock performance is in fact closer to that of a Cortex A8. I'd still argue its overall performance is high enough to be "competitive" with an A9 though, in the sense that they will provide very similar user experiences. Exynos may offer much greater peak performance with its four cores, but I'd bet typical core utilization is even lower than on the desktop, further narrowing the effective gap.

    Don't get me wrong, more performance is always nice, and it's unfortunate that Intel isn't introducing a new architecture until 2013, but given the relatively simplistic usage scenarios of current tablets, I have no doubt Intel will be ready in time for the real "post-PC" revolution. ;-)
    Reply
  • Wilco1 - Tuesday, October 02, 2012 - link

    Yes the micro architecture is nearly identical. However LPDDR2 is slower than DDR3 and I'm not sure whether all Z2460 based phones utilise the fastest memory speed. Also the Z2460 can only turbo up to 1.6GHz for a short time, unlike the N455 which can continuously run at 1.66GHz (but uses 6.5W). So a Z2460 is never going to be as fast as the N455.

    Note Z2760 results have now also appeared on Geekbench: http://browser.primatelabs.com/geekbench2/compare/...

    I agree performance seems good enough, but it just isn't better performance like Intel's marketing claims (which is how this thread started).
    Reply
  • tuxRoller - Tuesday, October 02, 2012 - link

    I don't what specint tests and GB doesn't release their code, so I don't know how they test.
    Assuming these numbers are accurate, I would have to agree that Krait is not a next gen arch. A 5% architectural advantage is more a bump than anything else.
    Assuming the cortex-A15's are all they are supposed to be, they will blow all the rest of these chips out of the water.
    The main advantage the Apple cores seem to have is that they are using lpddr-1066 memory (looking at the GB breakdowns, it is their massive memory bandwidth and streaming perf that gives them their advantage). That seems to be what accounts for the difference, but perhaps when AT does the review we'll here more.
    Reply
  • tuxRoller - Tuesday, October 02, 2012 - link

    Actually, the difference isn't 14% but only 5% at the same clock, for that benchmark.
    I wasn't aware of any benchmark where krait performed that poorly compared to cortex-A9.
    Sure, Intel may have picked this benchmark b/c it makes them look particularly good, but I'm more concerned than with the A9 vs krait numbers.
    A 5% architectural advantage seems absurd.
    Reply
  • Krysto - Friday, September 28, 2012 - link

    Clover Trail is still not competitive with ARM. It needs a 30 Whr battery, which is more than twice as large as most 10" tablets with a similar resolution (not counting iPad which has a much larger resolution. Clover Trail won't even support such resolutions, because they know battery will die too quickly).

    So even if they achieve the 10 hour mark, which I still doubt it. They're doing with a double battery. Doesn't exactly make for a "competitive" chip with ARM, does it?

    They're being misleading about a ton of other things, too. I suggest you give this a read:

    http://semiaccurate.com/2012/09/27/intels-clover-t...

    Also, it's a chip that's launching at the end of 2012. Where is the OpenGL ES 3.0 support?
    Reply
  • Death666Angel - Friday, September 28, 2012 - link

    No idea what you are talking about battery wise. Most 10.1 tablets I know of use ~25 Wh batteries ( Asus, Samsung, Acer, Chinese tablets). Reply

Log in

Don't have an account? Sign up now