Huawei’s Kirin 980 is the world’s first 7-nm mobile SoC

Apple and Qualcomm might get the lion's share of attention for their smartphone SoCs, but Huawei and its HiSilicon arm have long been making leading-edge mobile processors of their own. At IFA, Huawei revealed the HiSilicon Kirin 980, its next-generation flagship SoC.

Richard Yu, Huawei's Consumer Business Group CEO, introduces the Kirin 980

According to Anandtech's detailed report on the chip, Huawei describes the Kirin 980 as the first mobile SoC to be fabricated on TSMC's 7-nm process. For general-purpose computing, it has four of the latest high-performance ARM Cortex-A76 CPU cores inside, as well as four more Cortex-A55 low-power cores. Huawei organizes the Cortex A76es in an interesting fashion, though.

Instead of allowing all four of those A76 cores to run all-out, Anandtech notes that Huawei's architects subdivide the A76es into a pair of max-performance cores running at 2.6 GHz and a pair of midi-performance cores running at up to 1.92 GHz. Combined with the high-efficiency Cortex-A55 cores running at 1.8 GHz, that arrangement gives the Kirin 980 plenty of options to let programs run on the right set of cores to balance performance against efficiency.

Huawei's internal performance benchmarks suggest the Kirin 980's high-performance cores will turn in a Geekbench 4 single-core score of 3360, compared to 2452 for the Qualcomm Snapdragon 845 SoC that powers many of today's flagship Android devices. That's not quite enough to catch Apple's A11 Bionic, whose composite score of 4207 from publicly-submitted benchmarks is still likely to be the best in the industry, but it's a nice boost to see from a non-Apple flagship SoC. The Kirin 980 has an LPDDR4X memory controller capable of addressing 2133 MT/s RAM, another claimed first for this SoC.

For graphics power, the Kirin 980 implements ARM's Mali-G76 GPU block. The Mali-G76 doubles the number of SIMD lanes per execution unit, leading to a massive potential performance increase. Huawei claims a 46% performance improvement using an unspecified workload versus the Kirin 970, plus vastly better power efficiency—a 178% improvement for the Kirin 980 versus its predecessor.

Modern smartphone SoCs are also including more specialized functional units like AI acceleration engines, and the Kirin 980 is no exception. Huawei has included not one, but two neural-network accelerators for tasks like image classification, and the company claims it can handle up to 4,500 inferences per minute, or slightly less than twice that of the Snapdragon 845's performance on the same internal benchmark.

To talk to the outside world, the Kirin 980 includes a Cat.21 LTE modem with claimed download speeds of up to 1400 Mbps. The device also has a custom Wi-Fi chipset with support for 160-MHz 802.11ac channel widths, leading to a peak claimed speed of 1732 Mbps.

Huawei says the Kirin 980 will appear in its upcoming Mate 20 series of phones, launching October 16, 2018. We'll be curious to see what the company does with its slice of flagship silicon.

Comments closed
    • Shobai
    • 1 year ago

    [url=https://www.worldipreview.com/news/huawei-to-pay-panoptis-10-5m-for-patent-infringement-16550<]Huawei has been ordered to pay for 4G patent infringement[/url<]

    • End User
    • 1 year ago

    Odd that Huawei did not compare it to the single core performance of the 10 nm A11.

      • blastdoor
      • 1 year ago

      Ha!

      I have noticed a trend over the last year or two in which sites, for example Anandtech, have simply stopped mentioning the existence of apples SOCs when reviewing SOCs for android phones. Also, gone are the ‘deep dive’ reviews of new Apple SOCs, which I found very interesting.

      I suspect it is a combination of advertisers and a large chunk of the readership of these sites being on the same page – – fingers in ears, eyes closed, “La la La la la”

      • Pancake
      • 1 year ago

      Why? They’re not trying to sell chips to Apple. They’re in the Android market and hence all relevant comparisons would be to other Android devices. Also, Apple aren’t selling their chips to Android device manufacturers. Would be a hoot if they did though!

        • End User
        • 1 year ago

        [url<]https://9to5mac.com/2018/06/26/android-iphone-switchers/[/url<]

        • tipoo
        • 1 year ago

        Because they’re the current high watermark, and Android vendors are happy enough to take pot shots at Apple in every other aspect.

    • Shobai
    • 1 year ago

    Where is Arm up to with big.LITTLE, or whatever successor they’ve got? Is the entire pool of 8 cores available for running any given thread? Or do the slow cores still get parked in order to transition to faster cores?

      • Shobai
      • 1 year ago

      The A75 introduced DynamIQ to succeed from big.LITTLE; I would imagine that there are 3 clusters, one each for the two pairs of A76 and a final for the A55s.

      • tipoo
      • 1 year ago

      big.LITTLE moved to full HMP a while ago, as in all cores can be used at once if the situation calls for it. The governor would probably want to avoid that happening too much though.

        • blastdoor
        • 1 year ago

        I hate to think of what the Windows task scheduler would do with these heterogeneous designs. The spanking Linux delivered to Windows with the relatively mild heterogeneity of Threadripper 2990wx (not even heterogeneous cores, just heterogeneous access to memory) makes me think Microsoft was wise to abandon smartphones.

    • gerryg
    • 1 year ago

    Huawei is still facing huge questions about about the integrity of their products. A number of governments are concerned about hardware and/or software made by them that might allow China secret access to information/capabilities. You can google it. The fact there’s even a question makes me not want to trust it. Not sure any other vendor is trustworthy, but in most cases they would spy for financial gain, as opposed to espionage/intelligence gathering. Hard for the average consumer to know. But overall good reminder that any connected/connectable phone/device/computer, and associated services/clouds, comes with a degree of risk you should acknowledge, and, as needed, mitigate.

      • Arag0n
      • 1 year ago

      And you don’t feel it’s fishy that the country pushing for those “suspicious” behaviours is the same country that hosts all Huawei rival companies and fails to provide any single evidence about how it could be possible for Huawei to spy on behalf of Chinese government… right..?

      Because US always tells the truth, like doing PPT presentations in the UN about the WMD Iraq had and was going to use to destroy the rest of the world?

        • chuckula
        • 1 year ago

        [quote<]And you don't feel it's fishy that the country pushing for those "suspicious" behaviours is the same country that hosts all Huawei rival companies and fails to provide any single evidence about how it could be possible for Huawei to spy on behalf of Chinese government... right..? [/quote<] LEAVE LUXEMBOURG OUT OF THIS!

          • smilingcrow
          • 1 year ago

          and Ireland.

        • aspect
        • 1 year ago

        It wouldn’t surprise me if US intelligence agencies have some software or hardware backdoors, maybe no specifically for phones, but available as long as it’s under the excuse of something like the Patriot Act.

          • Anonymous Coward
          • 1 year ago

          I think it would be too difficult to keep a hardware backdoor secret. (Edit: except perhaps on a closed platform such as a phone.)

            • Shobai
            • 1 year ago

            Like the one affecting VIA C3 CPUs, which were released in 2001, that was discovered in 2018?

            Given how few devices this will ship in, not to mention how little longevity it will have, I don’t see it being particularly likely that such a backdoor would be found before the horse bolts.

            • gerryg
            • 1 year ago

            Depends on the device and how “black box” it is. IIRC they were first called out for high-end routers? As in ones that ISPs use. They were super cheap compared to Cisco and others, which was the draw. Scary.

          • blastdoor
          • 1 year ago

          When defending the worst people, it’s a common tactic to point out that the best people aren’t perfect and then to act like “goodness” is a binary variable.

          But things like “goodness,” “privacy,” “product integrity,” “security,” and “freedom” are not binary, they are continuous. It’s not a matter of existing or not, it’s a matter of degree, and the degree matters.

          When you make the perfect the enemy of the good you are working on behalf of the worst.

    • ronch
    • 1 year ago

    Are these scores comparable to Geekbench 4 scores obtained by x86 PC processors? I just checked the posted online scores for the FX-8350 and it seems the FX gets around 2,000 in single core and around 11,500 in multi-core. Hmm.

      • tipoo
      • 1 year ago

      The workloads are the exact same, the only thing is to look at the sub-tasks in case a few scores (like decryption) are dragging something up where it’s losing every other test. But that too is an outdated concern, as x86 chips started decryption in hardware like mobile, and sometimes have higher scores there.

      If that makes the FX look bad(er), now consider that A11 is twice as fast per core!

        • ronch
        • 1 year ago

        My phone has 4 x 1.3GHz Cortex A53 cores and single core score is 585. The FX-8350 gets around 2,000 and presumably runs at 4.2GHz when running one thread. So if you do the math, (4200 ÷ 1300) x 585 = 1,890. Guess the FX-8350 isn’t much better in terms of IPC compared to the A53. Or at least it seems that way if the scores are more or less comparable.

          • tipoo
          • 1 year ago

          The A53 has impressive IPC. It just can’t clock particularly high. Pretty typical pattern, see Jaguar too, short pipeline, does a fair bit per clock, but that stops it from clocking as high as higher end parts.

      • DavidC1
      • 1 year ago

      I don’t know. It seems questionable.

      If you ever run Geekbench, each subtest only takes a few second to run. It also scales very well with clock speed.

      Two things that that’s usually not the case with average benchmarks seen in review sites.

      • dragontamer5788
      • 1 year ago

      Yes, but the Geekbench 4 scores are a bit synthetic.

      IIRC, Linus Torvalds still complains that Geekbench (most noticably Geekbench 3) is too small and fits inside of L1 cache most of the time.

      So these tiny phone-SOCs can execute everything with their tiny 16kB of L1, without going to L2. A desktop processor on the other hand, deals with bigger problems and has 256kB of L2 cache and 8MB of L3 cache to handle problems of various sizes.

      Another issue is that Phone-SOCs can often times turbo the entire Geekbench run, since it completes so fast. Desktop benchmarks generally run long enough for the processor to heat up and test the overall cooling capability of the system (gotta test that overclock somehow!!)

      Still, quick-turbo then rest is a pattern in the phone world, where you have to decompress an HTML page, parse the content, deliver it to the user and then… go to sleep to save on battery life. So Geekbench is a decent test of that use-case.

        • tipoo
        • 1 year ago

        iirc the workload size was resolved with 4, and the work sizes are now the exact same between desktop and phones, unlike before where phones did less work because it would take so long to run. And some phones (well, the iPhone at least, not sure of every SoC out there) do have similar cache sizes above 8MB total to their desktop counterparts.

        Where I do agree is that modern hardware finishes each test way too fast, in seconds, so it’s all in boost states. Geekbench 5 should last at least several minutes.

    • Fursdon
    • 1 year ago

    [quote<]... and a pair of midi-performance cores...[/quote<] Sounds awesome.

      • highlandr
      • 1 year ago

      Pretty sharp comment.

      They shouldn’t have any treble powering everyday tasks.

      This new chip is music to my ears.

        • Srsly_Bro
        • 1 year ago

        Because are you all about the bass?

          • Pancake
          • 1 year ago

          White people turn up the treble, black people turn up the bass.

          – Sgt. SLICK

    • soccergenius
    • 1 year ago

    Honest question: if/when (according to rumors) Apple announces new iPhones with 7nm A12 chips on Sept 12, available presumably Sept 19, will Huawei still be “first” if they don’t ship until a full month later?

      • tipoo
      • 1 year ago

      That’s not infrequent, people saying they’re late by comparing announcement dates to actual shipments dates (which for iPhones and macs is usually very close to announcement time…Airpower aside).

      • cygnus1
      • 1 year ago

      Came to post basically the same thing. Not much about this SOC is first, and it’s definitely not going to be first shipped.

    • DancinJack
    • 1 year ago

    Lots of cool features. Wish we could get the competition in the US.

    Though I’m sure A12 will just stomp on this thing in most tasks, and will likely only have four cores max.

      • tipoo
      • 1 year ago

      Nah. A11 is already hexacore. Though A12 will probably keep that, and the only two high performance cores.

      A12X is going to be scary for the U series though.

        • DancinJack
        • 1 year ago

        Not really sure how I didn’t remember that…weird. Thanks for the reminder.

        • blastdoor
        • 1 year ago

        I’m going out on a limb to predict there will be no A12X. Note that there has been no A11X.

        Instead, I predict that the A12 will be used in the iPad Pro, just at a higher clock speed. I further predict that once people see the performance, nobody will complain.

          • tipoo
          • 1 year ago

          I dunno, the A10X is notable for having twice the memory bit width than the iPhone parts, even the A11.

          If it was a A12 with a 128 bit memory bus, it would still be significantly different than the A12, whatever they called it, and at that point with the added (physical) width they may be able to add more GPU cores etc as the X part always has to plug in die area.

          Also the iPad _____ is where they’re putting the non-X parts now, the X is one of the things differentiating the Pro.

            • blastdoor
            • 1 year ago

            All great points, but I’ll stick with my prediction nevertheless 🙂

            I’m basically predicting that the A12 will be a beast that is under clocked in the iPhone in order to meet the power/thermal constraints of the device and clocked higher in the iPad Pro. I’m predicting that the higher-clocked A12 will demolish the A10X it replaces.

            • tipoo
            • 1 year ago

            We will see!

            I guess I also just enjoy watching it go in real time from “huh, this could almost meet Core M in a few years” to “Huh…They matched Core M in a single year” to “Uhh, this is getting up there with the U series” and beyond, and an X version would help accelerate that even if the A12 unbound was amazing 😛

            The rumored 3 + 5 core X version would be frightful!

            • blastdoor
            • 1 year ago

            Why not just have the A12 be the 3+5 version, but at a lower clock speed?

            The fixed cost of taping out a new design on these new nodes is so high — I wonder if even Apple is thinking that it makes more sense to use the same design for all products, just clocked differently.

            • tipoo
            • 1 year ago

            I think it’s been because a 128 bit memory bus was too power consuming for a smaller phone so far.

            Looks like Longhorn has been poking around too – two parts, CPU cores codenamed Vortex

            [url<]https://cdn.wccftech.com/wp-content/uploads/2018/05/s_998bd99fb85d4e2a8c133668c8efdb45.jpg[/url<]

            • blastdoor
            • 1 year ago

            Oh well — so much for my uninformed speculation!

    • chuckula
    • 1 year ago

    PrincipalSkinner would be proud!

      • Shobai
      • 1 year ago

      Of all the articles for him to miss out on…

Pin It on Pinterest

Share This