Qualcomm Snapdragon X (Oryon)

Jump to latest Follow Reply
continuum

continuum

Ars Legatus Legionis
94,897
Moderator
  • #1
    www.anandtech.com

    The Qualcomm Snapdragon X Architecture Deep Dive: Getting To Know Oryon and Adreno X1

    www.anandtech.com www.anandtech.com


    We'll get actual reviews by end of the month but looks like technical details are starting to come out.

    Basics of the currently released Snapdragon X Elite and Snapdragon X Plus chips are 10 to 12 P-cores (Oryon) in 3x 4-core clusters, no E-cores (aka no smaller cores), ARM v8.7-A ISA, Hexagon NPU up to 45 TFLOPS (INT8), LPDDR5x-8448 memory, TSMC N4 process.

    I'm gonna defer more sophisticated commentary to those who are much more knowledgeable about CPUs than I am, but this looks like an actual serious non-Apple ARM processor aiming for competitive performance vs. Apple, AMD, and Intel's current processors, so am very curious how this goes. Some FPU limitations in this current design look pretty significant so if this makes it to a second generation that might be even broader in appeal.
     
    steelghost

    steelghost

    Ars Praefectus
    4,975
    Subscriptor++
  • #4
    fitten said:
    I've been eagerly watching these. I'm eager for a Linux box with one of these in it.
    Click to expand...
    What's the chance of Qualcomm opening the driver source for these though? I'm thinking the GPU in particular is unlikely, from what I have seen of their behaviour with Android support. Binary blob city. Otherwise I'm right with you, would be great to have another performant architecture for Linux.
     
    A

    Anonymous Chicken

    Ars Scholae Palatinae
    1,134
    Subscriptor
  • #5
    Splitting the cores into three clusters of 4 answers the question of how they will compete with Lunar Lake on low-demand efficiency. They even have an L2 shared in each 4-core cluster. This does not appear to be the best way to get maximum performance though, seeing how AMD & Intel choose to make large clusters of cores when they can. I wonder if next generation they'll have a big cluster and a little cluster (even if the core architectures are the same in each).
     
    • Like
    Reactions: continuum
    R

    redleader

    Ars Legatus Legionis
    35,019
  • #6
    Aeonsim said:
    Internal 128b only Neon implementation is a bit of a pity, though there maybe 4 copies of it per core? But then the Apple silicon (at least up to m2) had 128b Neon as well...
    Click to expand...
    It has FP16 support, so for at least some things it can match the FP throughput of AVX (which does not have FP16 support since it's ancient), although at much lower precision. Overall though that's probably reasonable in an ultra portable laptop. The number of applications that would ever get 256 bit vectorization in that niche is probably close to zero.
     
    rodalpho

    rodalpho

    Ars Praefectus
    3,391
  • #9
    My concern is Qualcomm sounds super confident in their hardware but Microsoft hasn't said much about the x86 translation other than "it's good now". They've lied about windows on ARM being "good now" on two separate occasions before so I'm at most cautiously optimistic. If not for all the OEM buy-in this go around I'd assume the worst.
     
    • Like
    Reactions: grstanford, rek and continuum
    continuum

    continuum

    Ars Legatus Legionis
    94,897
    Moderator
  • #10
    That's a huge part of why I'm more than happy to wait for actual reviews.

    As excited as I am about the Oryon core, with AMD Zen 5, Intel Lunar Lake, and now Qualcomm as a player in Windows laptop processor space, this is looking like the most competitive scene for CPUs in a long time for Windows systems. I have a bunch of refreshes waiting for fall/winter to see how everything shakes out once the last of the three (Intel Lunar Lake) shows up... and that extra bit of time ideally will give both Qualcomm/Microsoft/AMD to sort things out in all the new stuff being put to use (Snapdragon Oryon, Windows Prism, and AMD's usual teething issues).
     
    • Like
    Reactions: teleos and rodalpho
    rodalpho

    rodalpho

    Ars Praefectus
    3,391
  • #13
    Initial reviews are out and Qualcomm's numbers appear to have been accurate.

    Bit slower than the M3 and x86 CPUs on lightly threaded, beats the M3 in multi threaded due to having more cores, and the windows x86 translation stuff really no BS works, comparable to Apple's Rosetta 2 with an approx 10% performance hit.

    GPU is comparable to AMD/Intel too, 720 high or 1080p low, and nobody talking about compatibility issues that I've seen.

    And battery life is stellar, a bit under the M3 but streets ahead of AMD and Intel.
     
    • Like
    Reactions: fitten and continuum
    B

    Bad Monkey!

    Ars Legatus Legionis
    21,393
  • #15
    hobold said:
    Charlie, over on semiaccurate.com, has just published a very unfriendly piece about the 'x86 emulator in ARM windows. And he is deeply mistrusting MicroSoft's "Pluton" security module, giving fairly detailed reasons as to why. The CPU silicon itself is actually decent, but the software around it seems not to be there yet.
    Click to expand...

    I'm not sure why anyone pays any attention to anything that Charlie over on semiaccurate.com says these days.
     
    • Like
    Reactions: Demento and Semi On
    Semi On

    Semi On

    Senator
    89,415
    Subscriptor++
  • #22
    steelghost said:
    What's the chance of Qualcomm opening the driver source for these though? I'm thinking the GPU in particular is unlikely, from what I have seen of their behaviour with Android support. Binary blob city. Otherwise I'm right with you, would be great to have another performant architecture for Linux.
    Click to expand...

    www.qualcomm.com

    Upstreaming Linux kernel support for the Snapdragon X Elite

    Latest SoC is getting traction on Windows-based laptops, and we’re also consistently upstreaming to mainline Linux. See areas of focus and try our Debian installer.
    www.qualcomm.com www.qualcomm.com

    Made in Hurry said:
    How many of you are going to buy yourself one of these?
    Click to expand...

    Not myself, but my employer is making them the default for employees so I should have one as soon as I'm up for an upgrade.
     
    • Like
    Reactions: continuum and steelghost
    Semi On

    Semi On

    Senator
    89,415
    Subscriptor++
  • #24
    Aeonsim said:
    Is the battery life really that much better? A number of the reviews seem to show decent/good battery life but nothing drastically better than the competition. One review had ~15% higher than comparable x86 machines, another had it at the upper end of the pack but with a mix of other X86 machines with longer battery life.
    Click to expand...

    Make sure you're looking at numbers normalized for battery size. Some of the OEMs have HOBL targets for certain class of machines and have taken the opportunity to reduce battery sizes.
     
    I

    IceStorm

    Ars Legatus Legionis
    24,871
    Moderator
  • #25
    Dave2D reviewed the Asus Vivobook S15. Battery life information is on screen for all of about 8 seconds, so pause quick:

    View: https://youtu.be/rSx0WZfDbE0?t=295

    Vivobook S15 - 70 Wh
    Razer Blade 14 - 68 Wh
    Zepyrus G16 - 90 Wh
    M3 MacBook Air - 67 Wh

    The Vivobook S15 runs longer than the AMD and Intel options. The lower the load, the more it outperforms them, and it's not a small lead. Over an hour at medium load, and 1.5-2hrs at light load.

    The MBA is slightly ahead on all counts (17-32min) with a 3 Wh (4%) smaller battery.

    One other good thing is that in Sleep, the Vivobook S15 only loses 1-2% overnight, as opposed to around 10% on a x86 device.
     
    Last edited:
    rodalpho

    rodalpho

    Ars Praefectus
    3,391
  • #26
    Yeah it's genuinely impressive on battery life. Also the webcam is supposedly very good, which many reviews just gloss over.

    Having seen more early reviews, it seems Microsoft's claims of universal compatibility for x86 executables were a bit optimistic, with reports of many games and the epic game store not working. A lot of that is a problem with the programs themselves, but that's what universal compatibility is supposed to solve, right?

    It's a very good first gen product. But it is a first gen product.
     
    I

    IceStorm

    Ars Legatus Legionis
    24,871
    Moderator
  • #27
    rodalpho said:
    It's a very good first gen product. But it is a first gen product.
    Click to expand...
    At the start of the video, Dave2D addresses that.

    This is far from the first attempt at Windows on Arm. Dave stopped reviewing Windows on Arm devices because they were universally awful. He says he wouldn't waste his time or his viewer's time on the review unless he was happy with the device.
     
    theevilsharpie

    theevilsharpie

    Ars Scholae Palatinae
    1,199
    Subscriptor++
  • #28
    rodalpho said:
    It's a very good first gen product. But it is a first gen product.
    Click to expand...

    This is Qualcomm's third or fourth try at an Arm-based Windows machine. It's not a first-generation product by any means.

    This is why I roll my eyes at the people predicting the imminent death of x86.
     
    • Like
    Reactions: grstanford, timezon3 and charliebird
    I

    IceStorm

    Ars Legatus Legionis
    24,871
    Moderator
  • #30
    rodalpho said:
    It's the first serious swing at it
    Click to expand...
    Windows on ARM dates back to Windows RT (Windows 8). This is far from the first "serious swing" at Windows on ARM.

    They've been trying to make Windows on ARM work as a shipping product since 2012. It's been 14 years, and at least three serious previous attempts.
     
    • Like
    Reactions: grstanford and charliebird
    R

    redleader

    Ars Legatus Legionis
    35,019
  • #32
    rodalpho said:
    Having seen more early reviews, it seems Microsoft's claims of universal compatibility for x86 executables were a bit optimistic, with reports of many games and the epic game store not working. A lot of that is a problem with the programs themselves, but that's what universal compatibility is supposed to solve, right?
    Click to expand...
    That is the usual tradeoff you make with emulation. You can be fast or you can be accurate, but not both.
     
    T

    Thegn

    Ars Legatus Legionis
    13,121
    Subscriptor++
  • #34
    rodalpho said:
    Apple did a much better job. Even at the M1 release everything that wasn't a driver or particularly low level pretty much "just worked". There were exceptions, but I don't recall any major ones.
    Click to expand...
    Apple also had the advantage of completely controlling the platform and having made the architecture transition twice - from 68k to PPC, and from PPC to Intel. I imagine that the process of adapting the platform helped them design processes that made the transition easier.
     
    • Like
    Reactions: continuum
    A

    Anonymous Chicken

    Ars Scholae Palatinae
    1,134
    Subscriptor
  • #35
    Aeonsim said:
    Is the battery life really that much better? A number of the reviews seem to show decent/good battery life but nothing drastically better than the competition. One review had ~15% higher than comparable x86 machines, another had it at the upper end of the pack but with a mix of other X86 machines with longer battery life.
    Click to expand...
    Yeah that was my impression from Dave2D's review. He talks fast and flashes the graphs faster, but from the angle he showed it, and against only two comparison PC laptops, it seems surprisingly similar to Intel & AMD on both performance and battery.

    Separately, I was idly wondering if the choice of 3x 4-core clusters had more to do with the origin as a throughput-optimized server arch for heterogeneous cloud workloads, or with the desire to have low idle power.
     
    continuum

    continuum

    Ars Legatus Legionis
    94,897
    Moderator
  • #36
    Anonymous Chicken said:
    Separately, I was idly wondering if the choice of 3x 4-core clusters had more to do with the origin as a throughput-optimized server arch for heterogeneous cloud workloads, or with the desire to have low idle power.
    Click to expand...
    From the Anandtech article it looks like they can scale the clusters to up to 6 or 8 cores per cluster, so I'm guessing it was for lower power-- easier to have power-gate more transistors/have fewer active cores that way.
     
    • Like
    Reactions: Semi On
    continuum

    continuum

    Ars Legatus Legionis
    94,897
    Moderator
  • #38
    I would assume very possible.

    Structure is a bit different on AMD, but AMD is 1MB 8-way L2 per core, 32MB 16-way L3 per 8-core CCX, so with a big enough shared cache (be it L2 or L3) and enough bandwidth into/out of said cache... I wonder what the transistor trade-off cost calculation they made is. Obviously not just in terms of transistors but power...
     
    A

    Anonymous Chicken

    Ars Scholae Palatinae
    1,134
    Subscriptor
  • #39
    Recall how for 14th gen desktop chips, Intel decided their L2's needed to be larger so they could save power on the L3. That is perhaps an extreme case, but there must be some meaningful underlying cost to a large, high-performance shared cache. When that shared cache is L2, it must be especially expensive.

    I bet their next version will have a bigger cluster with shared L3, next to a cluster like now with a shared L2. A bit of the direction that Intel went with Lunar Lake.
     
    End_of_Eternity

    End_of_Eternity

    Wise, Aged Ars Veteran
    175
    Subscriptor
  • #40
    Worth reading this article from Semiaccurate:

    SemiAccurate: Qualcomm AI/Copilot PCs don't live up to the hype

    Particularly in context of some early (arguably disappointing) results from early X Elite reviews.

    And then after reading this, look at Ars's coverage of X Elite over the past ~9 months. Specifically these three articles:

    Qualcomm Snapdragon X Elite looks like the Windows world’s answer to Apple Silicon (October 2023)

    Qualcomm says lower-end Snapdragon X Plus chips can still outrun Apple’s M3 (April 2024)

    Is the Arm version of Windows ready for its close-up? (April 2024)
     
    Last edited:
    You must log in or register to reply here.
    Jump to latest Follow Reply
    Jump to latest Follow Reply