Mobile processor CPU core counts increase again

Discussion in 'Off Topic' started by iMacmatician, Apr 22, 2015.

?

Out of all the devices that you own, the one(s) with the most CPU cores is a…

  1. …watch

    10.0%
  2. …phone

    30.0%
  3. …tablet

    10.0%
  4. …laptop

    20.0%
  5. …desktop

    70.0%
  6. …server/workstation

    20.0%
  7. (other)

    20.0%
Multiple votes are allowed.
  1. iMacmatician

    iMacmatician Member

    Messages:
    310
    Likes Received:
    0
    Trophy Points:
    0
    For many years, we have seen 8 core CPUs on the desktop, and we have also seen mobile SoCs with 8 CPU cores. Taiwanese company MediaTek has made the inevitable jump past 8 with the 10 core MT6797, which may arrive late this year.

    Many mobile SoCs today use ARM's big.LITTLE technology, which combines two types of cores: big, powerful, and power-hungry cores with little cores that use a small amount of power. The idea is that the big cores are only activated if the SoC requires more performance than what the little cores can provide, which results in lower SoC power use. So a big core may be running a 3D game while the little cores may be working with the camera and many non-intensive apps. While the big cores can reduce their power usage for low-performance tasks, the LITTLE cores are actually more efficient in that area, hence the benefit to using LITTLE cores as much as possible.

    Mediatek has taken this approach one step further, at least in the hardware, and the MT6797 uses three types of cores.

    [​IMG]

    The A72 cores are called "huge" cores, the 2.0 GHz A53 are the "Medium" cores, and the 1.4 GHz A53 are the "TINY" cores. Note that the second set of A53 cores are optimized for low power usage much more than the first set. If the first set of A53's are clocked at 1.4 GHz, then they should be worse (in perf/W I think) than the second set of A53's, and the same if the second set of A53's are clocked at 2.0 GHz.

    Here's a car analogy (I don't think I've seen one of these in a while) and the obligatory benchmark slide.

    [​IMG]

    I think one of the biggest beneficiaries of these types of designs are the marketers, who can advertise ever-larger numbers of cores. Intel and Apple are two notable companies with the property that each processor from them contains only one type of CPU core, and they haven't been vocal about core counts. I recall reading something about Intel not liking big.LITTLE-type concepts at all*—we haven't seen any Intel CPUs with Core cores and Atom cores in the same chip.

    Sources: Hexus, ARM.
    * That wasn't always the case.
     
    Last edited: Apr 22, 2015
  2. Space_Oddity

    Space_Oddity The Shitstorm

    Messages:
    2,958
    Likes Received:
    43
    Trophy Points:
    0
    Food waste bin has most cores in it.
     
  3. ImSpartacus

    ImSpartacus nerf spec plz

    Messages:
    8,598
    Likes Received:
    7
    Trophy Points:
    0
    Is a "3 core" setup like that still stock big.LITTLE? I don't trust mediatek to do their own thing if big.LITTLE isn't available. ARM has had a couple years to tweak it and it wasn't perfect out of the gate.
     
    Last edited: Apr 22, 2015
  4. iMacmatician

    iMacmatician Member

    Messages:
    310
    Likes Received:
    0
    Trophy Points:
    0
    I'm not sure. There's no mention of software (at least in the report), which makes me wonder if MediaTek even has the software to match the hardware. We seem to only have their reassurance that the three cluster setup results in "reduced power consumption vs ARM big.LITTLE across all performance demand conditions."
     
  5. McGyver

    McGyver Experimental Pedagogue

    Messages:
    6,533
    Likes Received:
    31
    Trophy Points:
    0
    It's only a 10 core chip if all cores can have workloads at the same time.



    Here have some Spartacus-tier half knowledge:


    [​IMG]
     
  6. iMacmatician

    iMacmatician Member

    Messages:
    310
    Likes Received:
    0
    Trophy Points:
    0
    From AnandTech: "The Mobile CPU Core-Count Debate: Analyzing The Real World."

    This article attempts to answer the question of whether a large number of cores and big.LITTLE processing is a good idea for mobile devices.

    The phone used is the Samsung Galaxy S6.
    • Exynos 7420: 4x Cortex A57 at 2.1GHz, 4x Cortex A53 at 1.5GHz

    In Linux, the list of the processes that are on a core is known as the run-queue. The run-queue depth is a number that represents how loaded the core is. For example, a depth of 1 means that the core is 100% loaded, a depth of 0.5 means that the core is 50% loaded, and a depth of 2 means that the core's load is shared between 2 processes, where each one by itself would fully load the core.

    While loading the BBC front page using Chrome, the LITTLE cores are constantly loaded, and while the big cores don't see as much load, they also see a decent amount of activity.

    [​IMG]

    [​IMG]

    The charts below show the frequencies and power states of the CPU cores while running the same task as above. The x-axis represents time, and measurements are taken about every 200 milliseconds, so each x value represents the interval between two consecutive measurements.

    Since CPU frequencies can change much faster than that, the cumulative time that the cores spend on each frequency is used. In the frequency distribution charts, for each x value the y-axis represents the (stacked) proportion of time that the cores spend on each frequency. So during the first 200 milliseconds, the LITTLE CPU cores spent about half of the time at 1104 MHz, a quarter of the time at 400 MHz, and the rest of the time at in-between frequencies.

    The power state charts are constructed the same way, but with power states instead of frequencies.

    LITTLE cores frequency distribution chart
    LITTLE cores power state chart

    big cores frequency distribution chart
    big cores power state chart

    The little cores are often active and generally run at 900-1296 MHz (potential frequencies 400-1500 MHz). On the other hand, the big cores are less active and don't exceed 1200 MHz (potential frequencies 800-2100 MHz).

    That's important because many people perform a lot of browsing on a smartphone, so any performance/power improvements in this area are welcome. Cost is also relevant. In the Exynos 7420, the big core cluster is slightly over 3 times the area of the LITTLE core cluster, so those SoCs with 8 A7s or A53s can hold an advantage over SoCs with fewer small cores while being cheaper than the SoCs with a few big cores.

    Also, higher per-core performance may not be as great of an advantage as one might have previously assumed.

    Note that currently the most powerful stock core from ARM is the Cortex-A57, so anyone who wants more single-threaded performance using an ARM architecture core will have to design their own (e.g. Apple's Cyclone) or get a core from someone else who does. So for those who use stock ARM cores, the only way to increase performance from the A57 is to add more cores (increasing frequencies too high is problematic for efficiency) and use power management to reduce the power consumption of idle cores.

    What we don't have is a direct comparison of, say, 2x Cortex-A57 vs. 8x Cortex-A53, and that's something I'd like to see. But what can be said is that high core count mobile SoCs can be properly utilized in nontrivial situations.
     
  7. Beerdude26

    Beerdude26 OnThink(){ IsDownYet(); }

    Messages:
    7,243
    Likes Received:
    13
    Trophy Points:
    0
    MediaTek plays on the same level as Qualcomm. They have sourced experts from all over the globe to come work on their SoCs. I've seen how their SoCs improved by leaps and bounds. First the MT6589, which was very nice little powerful quad-core, then the MTK6592, which was a huge performance increase but a battery killer, then (my current daily driver) the MTK6752, which upped the performance increase even more but reduced battery usage. It truly is insane how efficient this SoC is. The Helio X10 and X20 continue this trend of more performance with less battery drain. Keep in mind that they're all doing this on 28nm HPM. The competition is already at 20nm and struggles to have the same efficiency as MediaTek.
     
  8. McGyver

    McGyver Experimental Pedagogue

    Messages:
    6,533
    Likes Received:
    31
    Trophy Points:
    0

Share This Page