06
May
Intel

Most smartphone manufacturers have chosen to ignore Intel's mobile offerings in favor of ARM chips, but Intel is hoping to change their minds with its latest microarchitecture. Today Intel unveiled Silvermont, which reportedly will result in new mobile chips with three times the performance of current-gen Intel Atom processors. Alternatively, Silvermont will enable Intel's next-gen Merrifield smartphone chips to achieve the same performance levels as Clover Trail+ with one-fifth of the power consumption.

silver

Want more specifics? Other advertised features of Silvermont include:

  • A new out-of-order execution engine enables best-in-class, single-threaded performance.
  • A new multi-core and system fabric architecture scalable up to eight cores and enabling greater performance for higher bandwidth, lower latency and more efficient out-of-order support for a more balanced and responsive system.
  • New IA instructions and technologies bringing enhanced performance, virtualization and security management capabilities to support a wide range of products. These instructions build on Intel’s existing support for 64-bit and the breadth of the IA software installed base.
  • Enhanced power management capabilities including a new intelligent burst technology, low-power C states and a wider dynamic range of operation taking advantage of Intel’s 3-D transistors. Intel® Burst Technology 2.0 support for single- and multi-core offers great responsiveness scaled for power efficiency.

Silvermont is actually the architecture of choice for Intel's entire lineup of next-gen low-power applications, including everything from smartphones and tablets to the next batch of Ultrabooks.

Of concern to us, though, is that the mobile chip platform which will be built on the Silvermont architecture, Merrifield, will begin shipping at the end of this year. Intel's current-gen Clover Trail+ processors haven't seen much in the way of high profile adoption, counting the likes of Lenovo and ZTE among its list of suitors. Meanwhile, Qualcomm has all but dominated the mobile chipset industry with its highly integrated Snapdragon SoCs. Samsung and NVIDIA continue to iterate impressive products, as well.

Based on this information, though, Intel is poised to be the first company to ship a mobile processor on a 22nm fabrication process, and those performance figures really do seem to mean business. Intel's previous-gen Z2460 chip, while no speed demon, was also far from slow, and actually demonstrated respectable power consumption figures. Now its next-gen successor will be more than three times as quick at full crank, and use less than a fifth the energy when completing more basic tasks. Those figures may have raised a few eyebrows at Qualcomm.

Source: Intel

Bertel King, Jr.
Born and raised in the rural South, Bertel knows what it's like to live without 4G LTE - or 3G, for that matter. The only things he likes sweeter than his tea are his gadgets, and while few objects burn more than a metal phone on a summer day, he prefers them that way anyway.

  • marcusmaximus04

    "three times the speed" != "3x higher performance". The first takes the form: 1x3=3. The second takes the form 1 + 3x1=4.

    • http://www.androidpolice.com/ David Ruddock

      I kindly disagree with that statement. Intel phrases it "3x more than," which translates in plain English to "three times more than," which is equivalent to greater than three times the existing level of performance - not 300% more. That would also be questionable marketing strategy - why on Earth would you call 400% of the performance a 300% increase?

      • marcusmaximus04

        re-read the statement "three times more than". You take three times the value and that's how much more power it has than it did originally. Hence, 1(the original amount) + 3x1(the amount more than it) = 4.

        This terminology difference is a major sticking point in computer architecture, and one I had slammed into my brain in college.

        Note that it's *possible* here that Intel used incorrect terminology. I'm just pointing out that the headline doesn't match what Intel is saying.

        • http://www.androidpolice.com/ Artem Russakovskii

          I'm also debating, but I think @rdr0b11:disqus is right regarding the 3x vs 4x here.

        • http://www.androidpolice.com/ David Ruddock

          I really think you're reading far too much into this. I've never seen anyone who does the math the way you're suggesting, it's hugely confusing, for reasons that are obvious at this point. I get what you're saying, I just don't think that's what Intel meant.

          • marcusmaximus04

            As I said, this is a very common issue in computer architecture(namely, the production of CPU's, which is exactly this context). It's something colleges focus on in intro to computer architecture: that "x time as much" is not the same as "x times more".

            You can set up a situation that makes this difference obvious: 1x as much as would clearly be the same amount. 1x more than would clearly be 2x the amount.

            Maybe Intel did misspeak in this case. All I'm saying is that some investigation may be warranted to check which one they mean.

          • http://www.androidpolice.com/ Artem Russakovskii

            Based on this slide and my consultation with Brian Klug from Anandtech, we've concluded that it's 3x total, not +3x, which means David Ruddock is correct.

            http://cl.ly/image/2J2L0k2E1h3u

            Looks like Intel took the best figure out of all the tests, which is 2.8x performance and rounded it up.

          • marcusmaximus04

            Thanks. Intel worded this in an extremely misleading way... But then from my experience, most manufacturers do the same with their benchmarks.

          • Dylan Patel

            Noone says three times and means 400%. Do people say double the speed and mean 300% of the original CPU? No noone does that. Not Intel, not Arm, not AMD, not Qualcomm. NOONE does that.

          • marcusmaximus04

            "Do people say double the speed and mean 300% of the original CPU"

            Of course not... Double the speed is the equivalent of 2x the speed. Now, if they said two times more speed, then yes, that means 300% of the original. That is literally exactly what that means.

            The thing is, most people don't even use that terminology outside lay people(along the lines of saying "I could care less"), or if they're trying to mislead.

    • ssj4Gogeta

      I agree with you that 3x != 3x more. But I'm guessing those slides weren't prepared by engineers, probably by marketing guys.
      Based on what I've read, it seems like it's 3x.

  • http://theotherlinh.com/ Linh

    The title is misleading, it's 3x peak performance OR 5x lower power consumption at the same performance, not both.

    • http://www.androidpolice.com/ David Ruddock

      It's clarified in the body of the text - what do you want us to title it - "Intel's New Silvermont Platform Delivers Three Times The Speed Of Existing Chips Or One Fifth Of The Power Consumption At Equivalent Performance Levels"? I think most people can manage to read the title, then the article, and understand what that means.

      • http://theotherlinh.com/ Linh

        A simple "or" could have been in place of the comma. Yes, the information is in the article, but it doesn't excuse a title that implies otherwise.

        • http://www.androidpolice.com/ David Ruddock

          I think you're reading into some kind of relationship between those variables that's not actually implied by the title. You assume that three times the performance and one fifth the power consumption means "three times the performance AT one fifth the power consumption," which obviously is not what it means. The relationship between the variables is not specified.

          Likewise, saying "Or" would potentially imply that you can have one in the chip, but not the other. You can have both in the same chip - just not at the same time.

          I've changed the title to "Peak Performance" to indicate more clearly what's meant here.

  • http://turbofool.com Jarrett Lennon Kaufman

    Historically Intel has appeared to be asleep at the switch for periods of time in which their competition have seemingly trounced them and taken over the market, only to have Intel surprise us by having spent all that "sleep" time developing new technology that not only retakes the performance crown, but retakes the market lead. This happened with them against AMD in desktop processors, and I suspect there's a chance it could happen here. The ARM architecture, and companies like Qualcomm, have a massive lead. But I wouldn't count out Intel's ability to prove the value of their architecture within the next few generations and force a transition. Hardly a done deal, but I don't think it's impossible.

  • VandrĂ© Brunazo

    What relationship does this have to do with Haswell? And is it a different architecture from x86, as in we would need to recompile any programs?

    • ssj4Gogeta

      The ISA is the same - x86. This just a different microarchitecture. Most Intel processors (Core series and Atom) are x86.

      This isn't really related to Haswell. This is an evolution of the Bonnell architecture of the current/original Atom, which has been unchanged since 2008 when it first debuted.

      • HopelesslyFaithful

        i don't really care about atoms until they include independent voltages...especially when it comes to small form factors

        • ssj4Gogeta

          You mean independent voltages for each core? With Silvermont each core has its own frequency plane, so maybe they have independent voltages too.

          • HopelesslyFaithful

            yes but that is a new feature with haswell. I am very excited to see it in action. My 3720qm uses 30w TDP (out of 45w) for a mere 25% utilization. I hope that independent voltages can get that down closer to 15w, which will allow much better battery life and better form factors

          • ssj4Gogeta

            If that workload was single-threaded, the other cores would be power-gated. Also, how are you measuring power consumption?

          • HopelesslyFaithful

            ThrottleStop. I have posts at Notebookreview about it. If one core needs 1.1v all cores are at 1.1v, which is where the huge power waste is. I also finished messing around with my 3720qm testing the scalability of the processor at various freqs and it is amazing how much juice is required to hit 3GHz and above. I can run at 2GHz across the board at full load using F@H and only use 17w, which is ~57% of total CPU capacity (out of 3.5GHz). But with running opera and moving tabs around and loading web pages I use 25% of the processor and use 30w TDP. 17w 57% vs 30w 25%. That is why i am stocked for independent voltages.

            Here are just a small tidbit of my results. With independent voltages you will be able to get a single core at 3.5GHz without wasting so much juice.

            This is F@H on all cores and you can see how high freqs require a lot of voltages.
            freq TDP
            3500 44.8
            3100 32.9
            2600 23.5
            2000 17
            1800 15.5
            1600 14.3
            1200 13

            using TS bench to just run 1 thread uses 12-20% of the processor and requires 21-23 TDP...that is basically half of the CPU TDP limit to run. Also windows or programs never run a program on one core even though it is single threaded. I honestly don't know why. I have tried setting affinity and it doesn't affect it. When i used my 920xm to play single threaded games i could never get it to turbo at single core speeds, which is why i have always been critical about the BS in turbo boost. I have never gotten my 3720qm to get to 4GHz either...always runs at 3.8-3.9GHz if i am lucky.

            here is an example. If you can explain why this happens i would love to know.
            http://img202.imageshack.us/img202/2803/tsbenchsinglethreaded.png

            Also on a side note can you please explain to me again how you would do the math for going from 22nm to 14nm in the increased efficiency. We talked about that before but i can't remember

          • ssj4Gogeta

            I also finished messing around with my 3720qm testing the scalability of
            the processor at various freqs and it is amazing how much juice is
            required to hit 3GHz and above.

            That's always the case. If you try overclocking a desktop processor, you'll notice that you hit diminishing returns after a while in terms of how much voltage it takes to get to higher clocks. Why do you think we moved from Netburst-like architectures (focusing on speed) to Conroe-like architectures (focusing on parallel execution and IPC)? :)

            With independent voltages you will be able to get a single core at 3.5GHz without wasting so much juice.

            Modern Intel processors have power gating, which can shut off a core entirely, giving you power gains more than any voltage scaling can. Of course, for that, that core has to be idle. A related Windows feature which makes use of it is "core parking".

            I'm not sure why Windows is migrating that thread between cores. The OS of course can do that and there can be many valid reasons for it. But migrating a thread encounters overhead.

            Also on a side note can you please explain to me again how you would do the math for going from 22nm to 14nm in the increased efficiency. We talked about that before but i can't remember

            Think of a square 22nm by 22nm. It has x number of transistors. When you move to 14nm, the square becomes 14nm by 14nm, but still keeps the x number of transistors. Gain in density is ratio of areas = (21*21)/(14*14) = 2.25
            "Full node" steps are chosen such that the next node is approximately previousNode/sqrt(2), which gives the new node around double the density.

          • HopelesslyFaithful

            yea but you forget that windows never allows a core to go idle...look at the chart. There is no reason a core with 1-3% at 3.8GHz has a load. All that can be thrown over to another core but since windows fails to do that it wastes a lot of electricity. I am interested with how this works with haswell. Each core will be able to scale much better as i started above. Again cores never turn off. I have never seen a core ever go to lowest state.

          • ssj4Gogeta

            yea but you forget that windows never allows a core to go idle...look at the chart.

            Note that the time needed to transition between states is of the order of 10's of microseconds, so it can be transitioning into and out of C6 state hundreds, or thousands of times. That graph won't reflect that, because it's not sampled at that rate. It just shows the average for, let's say, the previous half a second or so. However, there is a latency in transitioning between states which the OS should try to avoid.

            I tried running a simple single-threaded program on Linux, and it's scheduled on the same core for the entire duration. I'll reboot later and try it on Windows too.

            Also, the games that you mentioned, it's highly unlikely that they're completely single-threaded. The benchmark, of course, is probably single-threaded if it says so.

          • HopelesslyFaithful

            all those games are only single threaded. Source can be multithreaded or single threaded and in single thread it never turbos to single core speed. Plus Source single threaded is a joke.

          • HopelesslyFaithful

            oh also if let us say each core is going on and off into lowest C state many time a second should it report a a lower freq and not 3.8GHz 24/7? since it is an average? Also if i remember going into lowest C state takes a fair amount of time, which if i remember correctly from reading somewhere it is very unlikely it goes that low due to the latency for the core to wake up.

          • ssj4Gogeta

            each core is going on and off into lowest C state many time a second
            should it report a a lower freq and not 3.8GHz 24/7? since it is an
            average?

            I'm not sure that programs which report frequency usually work like that. I've never seen frequencies which correspond to weird multipliers, like 30.2 or 30.7. They always correspond to approximately whole numbers or whole number + 0.5.

            Also if i remember going into lowest C state takes a fair amount of
            time, which if i remember correctly from reading somewhere it is very
            unlikely it goes that low due to the latency for the core to wake up.

            That's correct, the deeper C-state you go into, generally more is the wakeup latency. The OS has to decide if it's worth it to go to a C-state. From this document:

            Even if it is possible to drop a core into a deeper CC-state, the OS has to ask itself various questions, such as what is the likelihood that processes are going to be doing more work very soon, so that dropping into a deeper CC-state might actually cost an unacceptable penalty? Similarly, the processor has to ask whether dropping a core into a lower CC-state is going to cause incorrect operation, say whether the delay in the processing of an interrupt will cause an event to be lost.

            all those games are only single threaded.

            You mean they do AI, playing background music, rendering, etc. all on a single thread? It's very highly unlikely that that's the case. Of course, the game can do majority of it's work on a single thread, and the other threads can be light on workload, but it will still require separate threads to do asynchronous tasks.

          • HopelesslyFaithful

            well i appreciate the help. I think with what i got so far i might write a well though-out and detailed email to unclewebb. Maybe he can help work out the rest of this. Thanks for the help.

          • ssj4Gogeta

            No problem. If you're on Linux, you can use "powertop" to check how much time your CPU spent in each C-state. It's like CPU Spy for Android if you're familiar with that.

          • HopelesslyFaithful

            If it comes to that I'll give it a shot.

  • EH101

    Oh how I miss the days of Hummingbird Exynos processors vs Snapdragon S2's and S3's vs OMAP3's and 4's... kinda.

  • http://twitter.com/Lammjr Lamm

    So, the entire android ecosystem is ARM based, thousands os apps and other stuff that depends on ARM instruction set and optimizations. Plus the freedom from manufacturers to choose chips from multiple sources and Intel expects everyone to move to single platform, with a completely diferent instruction set and a single source for chips, considering that intel does not license it`s X86 arctecture to anyone...

    Nah, I'm sticking with ARM on this.

    • Freak4Dell

      The number of apps that solely rely on ARM is much, much lower than you think. The only app I've come across so far that doesn't run on my RAZR i was Firefox. That had already been fixed in the nightlies at the time, and I just didn't know about it. Now, even the stable version runs fine. Chrome also had a problem, but that was fixed around the same time the RAZR i was released, so I didn't even see a problem by the time I got it. Other than those, everything I've tried has run just fine.

    • ssj4Gogeta

      and Intel expects everyone to move to single platform, with a completely diferent instruction set and a single source for chips

      A single source for chips that is also the biggest semiconductor manufacturer in the world and has the most advanced fabs. The desktop/laptop market has been doing just fine sourcing most of their chips from a single source.

      Well, at least in terms of having quantity issues. We definitely need competition to prevent prices from going up uncontrollably.

      Most of the applications can easily be recompiled or are Java.

      • http://twitter.com/Lammjr Lamm

        yeah, because all we need is another monopoly.

        I srsly doubt samsung would go for a Intel Chip, they would be shooting themselves on the foot.

        I don't see Intel becoming dominant in the mobile sector.

        • Dylan Patel

          Noone is talking about a monopoly. We are just saying Intel is bringing another option to the table.

          • hobagman

            No one is two words ... not one ... when you say "noone" you sound like a moron

    • Nickhe

      Well the Architecture doesnt matter that much as you think in the android acosystem . Because 95% Of apps are written in Java , Which is made to be a Crossplattform , multiarchitecture based language , so last time i checked 90% of the apps worked on Intel based android system without a struggle , the only problem is the compilers for the apps aren't X86 optimised so the apps are running in fullspeed ,if see it that way

      • ssj4Gogeta

        Compiler only matters for non-Java applications.

    • http://www.facebook.com/profile.php?id=1745689461 Hal Motley

      Despite being biased towards ARM. However Android is built in a way which makes it very compatible with other chipsets except at the lowest of levels. Using primarily the Linux kernel and Dalvik (which is Android's Java VM) ensures the platform is probably one of the most portable operating systems out there maybe even alongside NetBSD.

  • RBI411

    Well Motorola allegedly signed a long-term chip deal with Intel. Lets see if Intel can finally gets its chips placed in a flagship phone.

  • Sergii Pylypenko

    How many other companies do 22nm chips? Exynos 5 is 28nm. Would be great if Intel dumped x86 arch altogether, and use only x64, but that won't be happening anytime soon, at least not until Android becomes 64-bit (we'll have that in three-five years), and corporate people start using ARM or MIPS chips in their servers (five-ten years I guess).

    • HopelesslyFaithful

      ~14 months we will have 14nm :D

    • Dylan Patel

      When intel says x86, they mean x86-64. 64 bit. They are not doing 32 bit anymore.

      • Sergii Pylypenko

        They ARE including full support for 32-bit x86, and 16-bit x86 in ALL their processors, so you can STILL run 32-years-old MS-DOS on it, if you find compatible BIOS and harddrive. Whenever they decide to drop that support, we'll have all tech news speaking about that for a month straight.

  • http://twitter.com/BrotherJim3 Jim Hasak

    I think it's cool that an AP staff guy will respond to readers' queries. Thank you, David!

Quantcast