Apple's A15 Bionic announcements undersells improvements over A14

Posted:
in iPhone
Apple may be underselling the performance of its new A15 Bionic chip, with initial in-depth benchmarks and analysis of the iPhone 13 SoC pointing to it having considerable power compared to its rivals, and a hefty improvement on the A14.




During Apple's iPhone 13 launch, the company promoted the A15 in an unusual way, namely by comparing it against its competition rather than previous generations. Apple mentioned the A15 Bionic was 50% faster than the leading rival chips, with 30% better graphical performance for the four-core GPU, 50% for the five-core, but not how much it improved on the A14 Bionic.

An in-depth breakdown of the changes in the A15 by Anandtech reveals that there are considerable improvements in performance over the previous year's SoC design. This starts with Apple's use of two new CPU microarchitectures for its high-performance and high-efficiency CPU cores.

Just on the cores, the maximum frequency per core in the A15 reached 3,240Mhz, versus the 2,998 maximum of the A14. The report found an 8% improvement in single-core performance for the two-core cluster, but when both performance cores are running, there's a 10% average improvement.

The efficiency cores also saw performance improvements, going up 10.5% over the A15 with a maximum frequency of 2,016MHz vs 1,823MHz.

A change in the system cache has also helped performance, with Apple moving up from 16MB in the A13 and A14 to 32GB in the A15. Latency tests confirmed the system cache has increased, keeping memory accesses on the same silicon for longer instead of searching DRAM.

Apple didn't mention changes to the L2 cache of the performance cores, which increased by 50% to 12MB, matching the size of the M1's version. The access latency is also up from 16 cycles to 18, while the A15 is also capable of one-cycle accesses to cache lines in L1 cache versus three cycles on the A14.

The efficiency cores retain their 64KB L1D and 4MB shared L2 caches, but the L2 TLB has increased to 2,048 entries to cover up to 32MB. There's also faster DRAM access for the efficiency cores.

In performance tests, the A15 is seen to improve on various processing tasks "across the board" against the A14 for the performance cores, meanwhile power consumption for those same tasks on those cores are also lower. It is reckoned that the A15's performance cores improved energy efficiency by 17% on its peak performance versus the A14.

The efficiency cores also see performance improvements in their own right, with 28% increased speed while keeping efficiency flat.

"Compared to the competition, the A15 isn't 50% faster as Apple claims, but rather 62% faster," the report reasons. "While Apple's larger cores are more power-hungry, they're still a lot more energy-efficient."

Pre-release benchmarks using Geekbench pointed to a general 21% CPU gain in the A15 over the A14.

On the GPU side, the 3DMark Wild Life test saw the four-core GPU in the A15 improving 14% over the A14, while the five-core version managed a 30% improvement. At peak performance, the five core GPU is "essentially double that of its nearest competitor," with Apple "likely low-balling things again."

The improvements grew in the GFXBench Aztec High test, with a 46% improvement in the 5-core, 19% in the four-core.

Early benchmarks initially claimed the A13 would provide a 55% graphical improvement over the A14.

Despite the improvements, the report points out that there are still thermal limitations at play, with throttling impacting the A15 in the Pro models more than the standard iPhone 13. Anandtech suggests this could be down to a new PCB design that puts the SOC in the middle of the PCB "sandwich" rather than the outside, with the latter improving thermal transfers to a heat spreader.

In summarizing the changes, the report says the A15 isn't a "brute force iteration we've become used to from Apple in recent years," but it does include "substantial generational gains that allow it to be a notably better SoC than the A14."

Read on AppleInsider

Comments

  • Reply 1 of 13
    There are so many types of chip specs that can be touted, but ultimately it's the experience that these can deliver which is important.
    The A-series offers a wide range of additional processing capabilities which are not easily measured by benchmarking software. (e.g. enhancements to the AI capabilities).

    Apple can't go on stage and say the device is x% faster because individual results will vary by the developer's skill in leveraging the chip's capabilities. Apple can make broad statements about improved performance - and this is what they continue to do, they don't want to be bogged down in another meaningless "MHz race". It's not particularly useful to keep reminding the audience that the device is faster than before, that is a given and isn't a novelty worth spending much time on.

    Apple take a different approach with their pro-hardware because performance and capability are the reasons for purchase, here you will see a lot of detailed comparisons in commonly used pro titles, as well as details like compatibility with systems/racks/vesa mounts and so on.

    aderutterwatto_cobra
  • Reply 2 of 13
    elijahgelijahg Posts: 2,506member
    The CPUs are now so fast that the generational improvements aren't really noticeable in daily use anyway. If you were number crunching then a nn% speed increase actually makes a tangible difference, but an app launching in 0.51 seconds instead of 0.6 isn't perceivable.
    twokatmewFidonet127netrox
  • Reply 3 of 13
    Well, this bodes well for M2 and M2x iterations of the Mac SoCs ... higher efficiencies and larger caches mean you can stuff more cores in within a given thermal envelope.

    I suspect that M*x variants might shuttle GPU cores off into chipsets - we'll hopefully see soon - so unbounded GPU cores can achieve higher performance levels if physically spaced apart under some kind of heat spreader.

    Higher efficiency E-cores mean that you can easily double the number to handle more mundane tasks, and having no Qualcomm modems in there contributing to heating should also improve performance and thermal capacity.
    watto_cobra
  • Reply 4 of 13
    Apple trouncing ARM, Samsung and Qualcomm again, as usual. 
    williamlondonwatto_cobra
  • Reply 5 of 13
    hmlongcohmlongco Posts: 320member
    elijahg said:
    The CPUs are now so fast that the generational improvements aren't really noticeable in daily use anyway. If you were number crunching then a nn% speed increase actually makes a tangible difference, but an app launching in 0.51 seconds instead of 0.6 isn't perceivable.
    What do you think drives all of those new camera features like face tracking, portrait mode, cinematic mode, multi-shot HDR processing, and so on? Or on-device text dictation and processing? Face ID? ProMotion?

    It's the steady improvement in processing capability that enables complex software to do all of the things we've come to take for granted in real time with no lag.
    WgkruegerFidonet127chadbagmike1williamlondonwatto_cobra
  • Reply 6 of 13

    Apple trouncing ARM, Samsung and Qualcomm again, as usual. 
    ARM does not produce SOC's. Apple uses the ARM architecture but yes, they are trouncing Samsung, Qualcomm, and the rest.

    williamlondonwatto_cobra
  • Reply 7 of 13
    chadbagchadbag Posts: 1,521member

    Apple trouncing ARM, Samsung and Qualcomm again, as usual. 
    ARM does not produce SOC's. Apple uses the ARM architecture but yes, they are trouncing Samsung, Qualcomm, and the rest.

    ARM does provide a reference implementation design if I am not mistaken and many manufacturers base theirs off of this reference but Apple doesn't  (?? Am I remembering it correctly?)
    edited October 5 watto_cobra
  • Reply 8 of 13
    elijahgelijahg Posts: 2,506member
    hmlongco said:
    elijahg said:
    The CPUs are now so fast that the generational improvements aren't really noticeable in daily use anyway. If you were number crunching then a nn% speed increase actually makes a tangible difference, but an app launching in 0.51 seconds instead of 0.6 isn't perceivable.
    What do you think drives all of those new camera features like face tracking, portrait mode, cinematic mode, multi-shot HDR processing, and so on? Or on-device text dictation and processing? Face ID? ProMotion?

    It's the steady improvement in processing capability that enables complex software to do all of the things we've come to take for granted in real time with no lag.
    Not the speed of the CPU. Most of those use dedicated silicon, and those that don't do not need the faster CPUs, which is why a lot of the software exclusives on the new phones work just fine on older, jailbroken phones.

    Where did I refer to "processing capability"? I was talking about processing speed, not capability.
    muthuk_vanalingam
  • Reply 9 of 13
    chadbag said:

    Apple trouncing ARM, Samsung and Qualcomm again, as usual. 
    ARM does not produce SOC's. Apple uses the ARM architecture but yes, they are trouncing Samsung, Qualcomm, and the rest.

    ARM does provide a reference implementation design if I am not mistaken and many manufacturers base theirs off of this reference but Apple doesn't  (?? Am I remembering it correctly?)
    Yep.
    thtwatto_cobra
  • Reply 10 of 13

    Apple trouncing ARM, Samsung and Qualcomm again, as usual. 
    ARM does not produce SOC's. Apple uses the ARM architecture but yes, they are trouncing Samsung, Qualcomm, and the rest.


    ARM designs cores that others (Samsung, Qualcomm, Mediatek, Huawei….) use. Like the X1, A78, A55.

    Samsung have both tried their hand at making either custom cores, or modified designs based on ARM cores. They have failed and now are back to using ARM cores.
    williamlondonthtwatto_cobra
  • Reply 11 of 13
    KITAKITA Posts: 382member
    Some very impressive results for the A15's performance cores on TSMC's N5P. Arm (focus on their 2020 Cortex X1 in the S888 on Samsung 5LPE) is still not designing this caliber of core:



    GPU has some excellent peak performance, but it's clear the thermals cannot be sustained in an iPhone chassis:



    Taking into account the sustained performance, GPU performance might be an area that other manufactures can overtake Apple (AMD RDNA2 + Exynos?).

    I'd also note that the review didn't seem to touch upon the Neural Engine performance for AI, however, that's one area that Apple is rather behind on (just based off of raw numbers, ~50% the TOPS of the S888+).

    It's clear though that Apple's CPU lead is still very substantial with a lot of future promise as Apple moves to new designs on TSMC N4 (2022) and N3 (2023).


    muthuk_vanalingamAlex_V
  • Reply 12 of 13
    elijahgelijahg Posts: 2,506member
    KITA said:
    t seem to touch upon the Neural Engine performance for AI, however, that's one area that Apple is rather behind on (just based off of raw numbers, ~50% the TOPS of the S888+).

    I imagine that's because Apple's neural engine isn't as general purpose as the competitors', they have AI silicon specific to each application: things like portrait mode etc. So much like CPU speed it doesn't really matter what speed the generic AI hardware is since the application specific silicon is what is usually used. And comparing that to other CPUs is apples and oranges.
  • Reply 13 of 13
    techconctechconc Posts: 188member
    KITA said:
    I'd also note that the review didn't seem to touch upon the Neural Engine performance for AI, however, that's one area that Apple is rather behind on (just based off of raw numbers, ~50% the TOPS of the S888+).
    You had nice post until this part.  It seems you're a victim of Qualcomm marketing.  When comparing something like TOPS, you need to do an Apples to Apples comparison.  Qualcomm advertises TOPS for their chips as a COMBINATION of CPU + GPU + DSP.  Apple is only advertising the TOPS for their Neural Engine only.  Anandtech reviews of the SD865 and SD888 make note of this. 

    Example:  SD888+
    "32 TOPS AI
    (Total CPU+GPU+HVX+Tensor)"
    https://www.anandtech.com/show/16789/qualcomm-announces-snapdragon-888-plus-5g-speed-bin-at-3ghz

    Example: SD865
    "
    Qualcomm advertises 15 TOPS throughput for all computing blocks on the SoC and we estimate that the new Tensor cores roughly represent 10 TOPS out of that figure."
    http://https//www.anandtech.com/show/16789/qualcomm-announces-snapdragon-888-plus-5g-speed-bin-at-3ghz

    Finally, if you actually think Apple is behind with AI / ML performance, then you should take a look at actual benchmark scores which show just the opposite.  Geekbench now has a Geekbench ML test for example.  They show scores for CPU, GPU and NPU performance of the same test.  

    http://https//browser.geekbench.com/ml/v0/inference
    Hint: the A15 is ~ 70% faster for this type of work as compared to the SD888.
    muthuk_vanalingamthtwilliamlondonwatto_cobra
Sign In or Register to comment.