Qualcomm SoC & ARMv8 custom core discussions


How much of the boost for the Kryo over the A57's in the 810 is due to the process improvement vs micro-architecture? It seems like most of it comes from better clocks and improved throttling behavior. The most distinguishing advantages of Snapdragon on the SoC level vs its contemporaries seems to be its integrated on-die modem, and its custom GPU / DSPs now.

The memory score also seems like an outsized contributor to the 820's aggregate scores in many popular benchmarks and in other respects Kryo seems to be on part with the A72, so maybe an A72 w/ some memory controller tweaks is the way to go as was rumored for the CPU in next Exynos.
 
Last edited:
The memory score also seems like an outsized contributor to the 820's aggregate scores in many popular benchmarks and in other respects Kryo seems to be on part with the A72, so maybe an A72 w/ some memory controller tweaks is the way to go as was rumored for the CPU in next Exynos.
Indeed. Picking the best scores for Cortex-A72 and MSM8996 on Geekbench leads to this. Hard to say what the frequencies these chips are running at, but A72 definitely looks competitive, except for memory where MSM8996 looks very good.
 
Hard to look flipping back and forth. Anandtech Huawei Mate 8 review, the A72 stores higher on SPECint2000 than Kyro? Still reading.
 
Nice spreadsheet. Looks like we wouldn't be missing much (maybe we'd even get something better) if they replaced Kryo with a tweaked version of the A72. How apples to apples is SPEC2000? What possible performance enhancement / detriment would iOS or Android have relative to each other?
 
The Denver scores were on an older toolchain / flags so not directly comparable.
iOS scores are definitely not apples-to-apples as it's compiled on Apple's LLVM vs GCC on Android.
 
At least it's not a Java/Dalvik/ART app on Android... I'm guessing apple's LLVM is much more optimized than GCC on android in general? They have only a handful of targets to optimize for. Would it be possible to compile Spec2k using Google/Qualcomm's LLVM?
 
Cool, looking forward to seeing that. Also, for mobile chipsets that pack enough ram to run SPEC2k6...
 
I think it may be dual-sourced with GF.
“I do believe Samsung must use the 820 to be more competitive with Apple. 820 hits performance per watt levels Exynos just can’t hit,” Moorhead said.
That made me laugh. I think somebody might have been exposed to too much PR. Reminds me of something. Cough.
 
If you don't isolate the view to just the CPU in the 820, a wrinkle that has been smoothed over from the 810, and consider the entire SoC with its built-in modem, sat-nav, misc DSPs, and GPU over the Exynos what he says might very well be true. The 820 probably won't have a clear lead over the Exynos 9 in CPU performance per watt, but all it needed to do was address the major issues the 810 suffered from; I doubt Exynos 9 sports as much integration and allows for as simple PCB layout as the 820 though. I'm curious about what advancements in stacked RAM this next generation will bring too; perhaps the excellement memory performance from the 820 comes from applying TSV techniques instead of wire bonding.
 
Last edited:

"Qualcomm said it has always used a variety of fab suppliers and will continue to do so. "

I heard before that Qualcomm has already been fabbing with GF for a while but I don't think I saw hard evidence. This kind of bolsters that.

Snapdragon 820 being manufactured by Samsung doesn't mean by itself that Exynos won't be used in (all?) Galaxy S7, but it does mean Samsung has that much less fab capacity to makes Exynoses. So that could be a factor.
 
(Sorry long read/ramble:) ) Well my belief ( and many others on here) that benchmarks don't tell the full story in regards to real world performance have been justified with a few head to head 'speed tests'.
Over the last week or so I have seen a few videos of real world browser web page speed tests and app opening speed tests for all the main performance smartphone, 6s, mate 8, mi 5 prime, galaxy s7 ( both versions) and the results seem to go against what geekbench composite scores would suggest.

The only exception being iPhone 6s which having the highest single thread geekbench score and faster nand is still the fastest phone in day to day usage, On android however the geekbench single thread does not point to the fastest every day SOC, despite what many tech websites will have you believe.
It's generally believed in the main tech media that single thread geekbench and crappy composite benchmarks such as antutu point to the fastest SOC of a smartphone, some smarter websites point to screen resolution as being a factor, while others are on point enough to recognise that device specific optimisations are also a factor, especially browser and now in ever increasing regularity head dissipation, whether that be metal unibody designs, or even so called 'liquid cooling' ,as well as some web sites picking up on manufacturing process being a factor.

The best tech web sites go into better detail and look at a wider range of benchmarks, looking at nand performance as being a major factor in general smartphone speed, as well as deep diving on power consumption and efficiencies, such as the excellent Anandtech articles we see on a more regular basis.
How ever we are now seeing main tech web sites taking real world 'speed tests' and multitasking tests into greater significance, such as tech radar, although most people on here will rightly point out that smaller lone reviewers have been doing such head to head videos on YouTube for a few years now, such as phonebuff who I believe started this off.

Whilst I am fully aware that such tests are subjective in alot of cases, are hardly scientific such as humann error can be a factor, as well as many of the points mentioned above, I personally would rather buy my smartphone and choose the SOC for that smartphone based on these real world tests, however amateurish they may seem to certain tech purists, time and time again I have seen benchmark topping multi core smartphone being outperformed in day to scenarios which go against benchmark results and what tech web sites would have you believe is the 'fastest, most powerful bla bla) smartphone.

For instance, despite Snapdragon 820 beating the other SOCs in most performance metrics on paper, including most composite benchmarks and graphics, in head to head speed tests it is actually the slowest, and by quite a long way, here I will link a few videos which I hope to point this out:
Exynos 8890 beats it s820 when both versions of s7 are compared:
http://www.ibtimes.com/samsung-galaxy-s7s-exynos-8890-variant-beats-snapdragon-820-video-2337237

Just so we can remove any Samsung optimisation favouritism the xiaomi mi 5 prime also uses the Snapdragon 820 higher bin ( 2.15ghz) ufs 2.0 storage and 3gb ram :
This is a comparison between mi 5 (standard low bin 820 1.8ghz??) And last gen note 5 exynos 7420:

And the so called 'slowest' next gen SOC Kirin 950 in the mate 8 with the lowest geekbench single thread composite score and only eMMC 5.0 nand seems to be the fastest SOC in real world performance, at least when paired up with a 1080p display such as the mate 8.

As you can see despite the Snapdragon performing top of the lost in most standard android benchmarks real world day to day usage does not seem to correlate.
My reason for posting this long rambling post then is to gather some thoughts from the tech savvy peeps on here on why this is, what constitutes a smooth fast smartphone? Is multicore performance more important on android than what most people believe? Or is it my belief that single thread integer performance of a SOC\cpu being the most important alongside nand? Is kryo just a benchmark bully but falls short in real life usage? Or has the new architecture not being optimised properly?
Certainly anantechs article which showed kryo and krait having the same integer execution units raised my eye brows, is this what we are seeing here? Or is cortex a72 paired with TSMC 16nm FF plus just superior combination? Would then porting a benchmark such as spec INT 2006 to android now that we have 4gb ram be a better real world representative benchmark?

I just want to point out I think hi silicon and huawei have done an excellent job with the Kirin 950 and mate 8, which I feel is the best smartphone for most metrics out today.
Despite on paper having the 'slowest nand, slowest single core composite geekbench and slowest graphics, it's seems to handle games with its 1080p screen with aplomb likely in part due to the 16nm FF plus process, has brilliant battery life in part thanks to its 4000mah battery, and is the fastest phone through the UI and app/ web browsing amongst android phones, despite having a custom skin.

Again sorry for the ramble.. Discuss.

Edit.
 
Last edited:
That's funny, given the noise some sites made about how S7 with S820 was crushing S7 with Exynos, based on AnTuTu results. Forbes, will you publish some update now? :D

That being said, I wonder why S820 is slower. Drivers? Throttling? Some HW limitations?
 
Dunno in other phones but in s7, Samsung has promised to release FW update that fix the throttling.

There's also Sony Xperia m4 aqua that throttle very quick and Sony has promised for a fix but it still hasn't come.

What I don't get is.. How software update can fix Thermal throttling problem?

Does snapdragons have cTDP like Intel y series? So manufacturer can customize it however they want?
 
Could be many possibilities...such as Samsung optimising TW for their Exynos...Samsung first run of 820 is not perfect(explains why they were able to launch much faster than the rest, in bulk, but sacrifices minimum speeds)...maybe Samsung will patch 820...

In a review i saw, the reviewers G5's 820 was able to launch apps much faster
http://www.androidpolice.com/2016/0...an-iteration-but-one-samsung-can-be-proud-of/

On the other hand...i wonder if Samsung can also patch the Mali performance in 8890 to get up to speed with A530....Did Adreno team really created a beast of a mGPU as Antutu scores have proven?
 
Back
Top