Qualcomm debuts Kryo custom CPU for the Snapdragon 820

Qualcomm's Snapdragon 820 SoC is an ambitious one. We already know that the company is targeting 40% improvements for graphics performance and power consumption with the chip's Adreno 530 GPU. Today, the company is throwing us another bone with details of the chip's custom CPU block, called Kryo.

Kryo is a 64-bit, quad-core processor built on a 14-nm FinFET process, and Qualcomm says it'll be able to reach speeds up to 2.2GHz. What's more, Kryo is supposed to deliver twice the performance of the Snapdragon 810 with twice the power efficiency. For context, the much-maligned Snapdragon 810 uses off-the-shelf ARM Cortex A57 and A53 cores in a big.LITTLE config, so the return to custom-designed cores could be a big deal for the 820 series.

Those huge purported gains in performance and power efficiency might come not just from custom CPU cores, but also from an emphasis on "heterogeneous computing" for the whole SoC. Qualcomm describes this approach as "[combining] different functional cores of the system-on-chip (SoC), like the CPU, GPU and DSP cores, to achieve previously unattainable performance and power savings, rather than using the same core for different tasks."

To that end, Qualcomm is also introducing its Symphony System Manager block, which is supposed to help the Snapdragon 820 run tasks on the best-suited processor core or split a task among different cores to achieve the best possible performance and power efficiency.

Qualcomm says the reveal of the Kryo CPU is the last major piece of the Snapdragon 820 puzzle. There's no way to verify the company's eyebrow-raising performance claims without hardware in our hands, but the bar has certainly been set high.

Comments closed
    • ronch
    • 7 years ago

    Too bad Skybridge has been canned. It would’ve been interesting to see big core ARM CPUs making it to the desktop. Not that they’d stand a good chance to unseat x86, of course.

    • NeelyCam
    • 7 years ago

    Going from a leaky planar 20nm process that produced the 810 nuclear reactor to a 14nm FinFET process could bring power efficiency improvements in excess of 50%. I remember Intel’s claim that FinFET gave them 35% power consumption reduction for the same performance. 1/0.65=1.54

    • NeelyCam
    • 7 years ago

    [url=https://techreport.com/news/28970/qualcomm-debuts-kryo-custom-cpu-for-the-snapdragon-820<]From TR[/url<]: "Kryo is a 64-bit, quad-core processor built on a 14-nm FinFET process"

    • willmore
    • 7 years ago

    I’ll bet you a donut that not all of those performance and efficiency claims happen at the same clock/power settings nor workloads.

    For example it might be 2x as fast as the 810 at full clocks while it may not be anywhere near 2x as power efficient *at that clock*. It might be 2x as power efficient at lower clock settings.

    • auxy
    • 7 years ago

    It wasn’t a direct quote. That was really more tangential to my point anyway. (´Д⊂ヽ

    It isn’t a matter of the hardware offloading engine “suddenly not working in 2 years” — you’re being a bit literal with me here — but rather a comment on the fact that fixed-function dedicated units like all these DSPs they’re using require very specific circumstances: specific input formats for the data to be processed, specific operating system environments for the drivers that are available, and so on.

    It’s not like you’ll be able to throw any random application on here and get good performance. This is my point.

    • culotso
    • 7 years ago

    Please AMD, bring back consumer-level dual CPUs. Those were my last big wow. Athlon MPs were so wicked nice.

    • fyo
    • 7 years ago

    How will the Kryo compare to the A72?

    Arm already took the A57 and optimized it for FinFET, claiming an increase in performance of 15-50% (lowest in integer, higher in floating point, highest in memory) while decreasing the die size 10% on similar process and reducing power consumption.

    The Kryo is going to be up against the A72 when trying to get design wins, not the A57. With similar claims of performance improvements from their respective marketing departments, it’s anyone’s guess who will come out on top.

    Qualcomm certainly has room to beat the A72 in terms of performance, with the latter very much constrained by die size. Arm’s decision to keep the L1 cache small (48+32) and the decode width narrow (3-wide), and not support L3 cache, means there is a lot of headroom for per-clock improvements if Qualcomm is willing to sacrifice the die area. Going 6-wide and 64+64 L1 cache would, judging by the results Apple obtained with the A8, increase performance substantially. Whether feeding that will require more L2 cache, or some (cheaper but slower) L3 cache, is a good question, though.

    • Welch
    • 7 years ago

    Depends if the claim if that their new core is 40% more efficient or that the 820 is 40% more efficient in general. If the former, then it may be the fuzziness your talking about.

    • fyo
    • 7 years ago

    Couldn’t find the quote you lambast them for and Google returns only this TR article with your comment. Not saying you made it up, just couldn’t find the source. Was it spoken as opposed to an official press release?

    On a different point, why would the hardware offloading engine suddenly stop working in 2 years? I doubt Qualcomm will introduce extensions to ARM v8, so everything should keep working just fine, at least until workloads change substantially — by which point the product shipping with the chip is going to be obsolete anyway.

    • Meadows
    • 7 years ago

    My first actual 3D accelerator had been a Monster as well. It’s a shame that nowadays the name evokes nothing but poopy cables.

    • Meadows
    • 7 years ago

    I was rather reminded of ye olde Hercules Kyro, but that’s because my long-term memory seemed to be dyslectic and I had to look up whether it was spelt “Kryo” or not.

    • NoOne ButMe
    • 7 years ago

    depends where you clock it, given you find the place on the FinFETs where C doesn’t increase to much and voltage decreases quite a bit you can get some pretty good gains.

    • auxy
    • 7 years ago

    This is like, Rockchip or Mediatek or even Allwinner-tier marketing. NO disrespect intended to the engineers at these companies who work hard and produce capable products, but their companies’ marketing departments need to be shot in the head.

    “Super quad-core processing for excellent execution of every software!”

    I’m not even making fun of their English skills but just the way they make ridiculous claims and emphasize things that don’t even matter. It’s really transparent to even the most casually informed user. I don’t know why they even bother; it just makes them look bad.

    MSI, Asrock, and the other smaller motherboard companies do this too. I really wish ANY of these companies would contact me and offer me a job writing their marketing copy. I could do a better job than the people they have for half the price. Seriously. I even read a little Chinese.

    Anyway, the claims made here are outrageous and I’m sure as you said they’re counting 4x efficiency as “we offloaded stuff to fixed-function hardware that will be poorly supported in 2 years!” Ugh. (´Д⊂ヽ

    • tuxroller
    • 7 years ago

    At this point, the die shrink can not bring anywhere close to that kind of efficiency gain. Maybe 20%?

    • blastdoor
    • 7 years ago

    Not a lot of details. Being way better than the worst thing they ever made is kind of vague.

    • Voldenuit
    • 7 years ago

    My first 3d accelerator was an S3 Virge.

    Feel my pain.

    Good thing Monster 3D came to the rescue.

    Going even further back, I have fond memories of my Tseng ET6000 bit blitters.

    • bthylafh
    • 7 years ago

    It’s going to have great benchmarks in Quake 3, just you wait!

    • wingless
    • 7 years ago

    I hope this chip will be what the Pentium was to the 486. I want to be wowed again…..a feeling I haven’t felt since the Athlon X2 dropped.

    • rika13
    • 7 years ago

    My first GPU was a Kyro and I thought they were naming it Kyro, not Kryo. The more interesting thing is that the GPUs in the Apple chips are descendants of the Kyro.

    • Voldenuit
    • 7 years ago

    [url=http://arstechnica.com/gadgets/2015/09/snapdragon-820s-custom-cpu-is-twice-as-fast-efficient-as-disappointing-810/<]From ars[/url<]: "the 820 will be built on a 14nm FinFET manufacturing process"

    • ozzuneoj
    • 7 years ago

    I thought the same thing!

    • Voldenuit
    • 7 years ago

    Kryo, the much anticipated successor to [s<]Pyro[/s<], I mean, 810!

    • DPete27
    • 7 years ago

    Anyone else find it funny they named it “Kryo” after the 810 had overheating problems?

    • jihadjoe
    • 7 years ago

    I wonder what their next CPU will be called. Kratos perhaps?

    • Peldor
    • 7 years ago

    820 is a 1% increase over 810. Surely such a revolutionary chip with a completely new architecture deserved a bigger number?

    • tipoo
    • 7 years ago

    Ah. So, a bit of fuzziness is the best guess. “We’re jumping a process node, and with finfets, and oh look how much more efficient our new core is!”. I wonder how the new core would compare to the old, like-for-like, on the same process.

    • chuckula
    • 7 years ago

    If the rumors are correct then these chips are definitely made with finfets, which is pretty crucial considering the major issues they ran into with the 810 on the 20nm planar process.

    • Jamannetje
    • 7 years ago

    I’m not sure that’s a good name for CPU, makes me think of PowerVR

    • tsk
    • 7 years ago

    This will likely be in my next phone once my Nexus 5 dies. Very pleased with the Snapdragon 800 performance.

    • tipoo
    • 7 years ago

    Now is this efficiency increase on the same manufacturing process, or including a die shrink? Because the Exynos in the S6 uses A57s/A53s to good effect, thanks to Samsung being ahead of other mobile SoCs with their fabrication plant.

    If it’s including the die shrink for the efficiency gains, that’s a bit of fuzzy marketing.

Pin It on Pinterest

Share This

Share this post with your friends!