Sony Xperia XZ4 with Snapdragon 855 is now on Geekbench too
- ?
- Anonymous
- tDR
- 04 Jan 2019
dude111, 04 Jan 2019Unless QC knows black magic i doubt their implementation ca... moreIf you read abit about micro architecture, tiny modification on cache can greatly affect performance and power usage.
I haven't read newer one but in older kyro, iirc they modified / reduce cache size to reduce power consumption and retain performance on most app (some app got penalties), resulting higher top clock.
Placement layouts of cores and io can also greatly affect latency and heat dissipation resulting in higher sustainable performance
- l
- lol
- SXy
- 04 Jan 2019
J o m s, 04 Jan 2019TLTR (Too Long To Reply) is that right?
Claiming A12X is... moreIt's a benchmark dude, its purpose is to compare two devices on a set list of tasks.
Do you even operating systems 101? Geekbench does not control the process affinity, no software does, it's all up to the scheduler.
If you have a rooted Android device, change the processor scheduler from schedutil to performance to interactive to on-demand and watch how the score changes accordingly.
Please take a look at the Phoronix review of the Threadripper and its followup against the Skylake-X and the review goes in detail how much faster multi core processors are in benchmark on newer Linux kernels than on Windows because of it being optimized for highly threaded hardware.
All Geekbench does is measure how much your device completes a given set of tasks. The fact that you set the affinity yourself means that you have overridden the default scheduler. If you want to blame anyone, blame Microsoft. Your experiment only proves that Geekbench properly responds to changes in the OS - Windows didn't give it enough resources so it only scored below 30K, you did it manually it went above 30K.
You want to compare it in a different set of tasks? Run different benchmark! Run PCMark or SPECperf, and in similar tasks the results are the same.
Just please read some actual analyses before spewing misinformation or take some computer science classes. Anandtech's deep dive into A12 is a good start, easy to find via Google.
And yes, A12 is getting dangerously close to the desktop level. It is the result of dropping largely compatibility-focused instruction set, smaller node, brilliant design and simple power electronics. There are many analyses on this topic and multiple benchmark suites not just GB that all support this conclusion. Unless you have a better argument than "whoah no way small chip more good than big chip!" just concede already and save yourself the embarassment. Benchmarks are more than just a number and you need to understand how to interpret them.
- J
- J o m s
- k1G
- 04 Jan 2019
dude111, 04 Jan 2019From cinebench: Rizen 3 1200 4 cores, 4 threads single ... moreIt does but not with single core score obviously. The.multicore ratio is lower than 1 but it does scale with each additional core.
- J
- J o m s
- k1G
- 04 Jan 2019
PCGuy, 04 Jan 2019Your experiment produces valid data but you fail to properl... moreTLTR (Too Long To Reply) is that right?
Claiming A12X is faster or as fast than Skylake or Coffeelake takes a lot of proofs to be true. So far only in GB and particularly with tasks optimized for ARM features so yeah GB is an invalid tool for comparison. You compare A12X with A12 or A11 within the same iOS version. Not with Android or Windows or Linux.
And you said it yourself
"Geekbench is not meant to be highly parallel benchmark, some drop-off is expected"
Why would I limit the encoding or compression of multiple files using poorly coded tool when I can assist it to maximize its potential?
"Running two instances and adding up is not even cheating, it's just invalid measurement method."
Nope. That is perfectly valid measurement. Lets see if.A12X can perform those experiments and claim that it is faster than Ryzen or Coffelake.
I was able to finish two tasks at the same time a GB does to the processor with only one task so yeah measurement is valid. Instead of GB forcing a task to run in multiple threads, it better spawn multiple tasks and wait for all of them to complete. Their you can tell which processor is better in multitasking. OS is there to help, not to be ignored by one incompetent benchmark claiming it is multiplatform. =D
And yes, I know the arch of ARM and X86 so I did not say SD855 should scale linearly in GB.
- B
- Bewildered
- 39x
- 04 Jan 2019
Nirvan2301, 03 Jan 2019I still have some doubts about the XZ4. Sony tends to be qu... moreThe Snapdragon 855 is 5G just like the Kirin 980.
The A12 Bionic is 4G as I understand it
The Qualcomm Apple court battle is in part over Qualcomm patented technology.
Qualcomm has if memory serves lost a recent court case and will have to license this technology.
This will reduce per unit prices Qualcomm charges phone manufacturers, Apple will like All benefit this, however Qualcomm in Germany and China have lost.
The 2019 iPhone Imo have to be 5G so Apple need to install the Snapdragon 855 no ifs or buts
History has shown me that Apple at this strategic point won't significantly increase the 2019 specs https://www.gsmarena.com/compare.php3?&idPhone1=4910&idPhone2=6378&idPhone3=7242
The IPhone 6 4.7 inch display is simply the 4 inch iPhone 5 with the non phone related Apple Pay, anyone on a 2 year contract iPhone 5 in the main upgrades to a 4.7 inch iPhone 5.
Samsung Galaxy 10 models will be both 4G Snapdragon 845 and 5G snapdragon 855.
- d
- dude111
- 8{c
- 04 Jan 2019
J o m s, 04 Jan 2019Try Cinebench or Blender... GB is a toy benchmark optimized... moreFrom cinebench:
Rizen 3 1200 4 cores, 4 threads
single 132
multi 482
It does not scale linearly either...
- d
- dude111
- 8{c
- 04 Jan 2019
AnonD-558092, 04 Jan 2019Maybe the K980 with the stock A76 can't sustain this perfor... moreUnless QC knows black magic i doubt their implementation can consume less, cause again, its clocked higher.
Yes, Huawei cheated before, maybe even now, thats no doubt, but if they achieve just that through cheating its highly doubful that the s855 can sustain those clocks beyond one or two runs. The A76 is much more power hungry than the previous A75 and A73 even with the node process at his favor.
- J
- J o m s
- k1G
- 04 Jan 2019
dude111, 04 Jan 2019lol There is basically no workload that can 100% paralle... moreTry Cinebench or Blender... GB is a toy benchmark optimized for iPhone.
- O
- Overclock.net lurker
- tDR
- 04 Jan 2019
J o m s, 03 Jan 2019So I did an experiment to prove that GB is not a proper ben... moreWhether GB is properly coded or not. You should know more core used = lower boost clock. Also, when more core used they prone to bottleneck other part, especially memory. High latency from infinity fabric also reduce overall multi core performance.
For qualcomm (or any SoC), we know on top they got strict thermal limit and low memory bandwidth, they also use multiple Core design with multiple speed inside a single SoC, unless they put homogen core on testing board with sufficient colling and memory bandwidth, there will be no anywhere near perfect scaling
- D
- AnonD-558092
- 45}
- 04 Jan 2019
dude111, 04 Jan 2019Well yeah, it should beat it as its clocked higher... But... moreMaybe the K980 with the stock A76 can't sustain this performance (well-known bench cheat...), and I highly doubt Sony has decided to cheat benchmarks.
- d
- dude111
- 8{c
- 04 Jan 2019
AnonD-558092, 04 Jan 2019It still beats the Kirin 980, which is exclusive ARM. The K... moreWell yeah, it should beat it as its clocked higher...
But at the same clock i dont think any benchmark would notice the "QC improvements" as realistically speaking, those kryo XXX cores should be 90-95% ARM, 5-10% QC tweaks
In fact i think that at the same clocks, the kirin should score a bit higher as it have more cache overall.
- F
- Furiounx
- P$W
- 04 Jan 2019
Walter C. Dornez, 03 Jan 2019Doesn't it keep the phone fast though?Well via hardware: yes. But you'd have to hope the software updates keep it fast too.
- D
- AnonD-558092
- 45}
- 04 Jan 2019
dude111, 04 Jan 2019GB only reflects the CPU part... which is ARM made.It still beats the Kirin 980, which is exclusive ARM. The Kyro cores are good custom cores. Not built from the ground up, but it's more than just ARM.
Either way, even if it was only ARM, Sony would have nothing to do with it. ARM Holdings does ARM designs on their own.
- P
- PCGuy
- SXy
- 04 Jan 2019
J o m s, 03 Jan 2019So I did an experiment to prove that GB is not a proper ben... moreYour experiment produces valid data but you fail to properly analyze because you lack the knowledge on how the hardware you're using and the benchmark works.
1. Geekbench is a perfectly valid benchmark and yes Apple's A12X really is that fast (AnandTech's deep dive into A12 architecture is a great read). The ancient compatibility-first, performance- second x86-64 instruction set your desktop is using is nowhere near as efficient as ARM modern short instruction word FOR THE THINGS THAT GEEKBENCH TESTS FOR. Usain Bolt is fast but you wouldn't expect him to outswim Michael Phelps. ARM has hardware acceleration for many of those things, and yes for some of those particular tasks A12X is faster than your Ryzen. Throw some highly parallel tile-based rendering and A12X will cry while your Ryzen will fly through it like a jet plane.
2. Your desktop may use 10x the power and require 10x the cooling but is not 10x faster. Smartphones fall on the very highest point of the (almost) exponential power-frequency curve while desktops fall at the very bottom. Try increasing the frequency from 4.2 to 4.3GHz and you'll see maybe 2% increase in perf but a much higher increase in voltage will be necessary so the power requirements will increase doubly since you'll need to supply more power as is and also due to the capacity of the material to conduct it. GamersNexus has a good article on it explaining the frequency-voltage curve.
3. Your Ryzen is on 12nm aka 14nm+ node (equivalent to Intel's 14nm). A12X and S855 is on 7nm node (equivalent to Intel's 10nm). That makes the point 2 even more important, since the 5-10W A12X will get much closer to the 95-150W 2700X.
4. Geekbench is well coded, what your experiment proved is the issue with Windows scheduler that is a known issue on Ryzens (and also higher threaded Intels). What you did by setting the affinity is launch two instances of Geekbench on two separate CCXs on the Zen die - LITERALLY THE BEST CASE SCENARIO FOR RYZEN. Geekbench does not control the process affinity, Windows does! If you want to blame someone, blame Microsoft but spoiler alert: this particular instance is the expected behaviour. Geekbench is not meant to be highly parallel benchmark, some drop-off is expected. Running two instances and adding up is not even cheating, it's just invalid measurement method.
5. Snapdragon 855's leaked score is perfectly in line with what Qualcomm revealed at their recent summit, if not a little too high. The single score is 3500 because of the one large special Kryo 485 Gold core that Qualcomm made a wider pipeline for and gave it twice the cache while probably giving it a higher power limit. The other 3 cores are much narrower Kryo 485 Golds at lower frequency and 4 even smaller Kryo 485 Silvers at even lower frequencies. You can't just 8 x 3500 dude, the other 7 cores are nowhere as fast. Even then, most tasks don't scale linearly - you can have 1000 cooks in your kitchen but you can't start baking a cake before you make the dough. Only in some very special instances like making 1000 cakes at once will having 1000 cooks make sense - these are called highly parallel tasks.
No, the Geekbench is not flawed and is a perfectly valid benchmark and you're insulting people who know much more about it because you don't understand the principles behind it. Feel free to contact them at Primate Labs maybe they'll explain it properly or take some CS classes but don't go about insulting people in the industry with your elementary school level experiments.
- d
- dude111
- 8{c
- 04 Jan 2019
AnonD-558092, 04 Jan 2019QC did the GPU, the DSP, the modem... We are past ARM designs.GB only reflects the CPU part... which is ARM made.
- D
- AnonD-558092
- 45}
- 04 Jan 2019
dude111, 03 Jan 2019*ARMQC did the GPU, the DSP, the modem... We are past ARM designs.
- d
- dude111
- 8{c
- 04 Jan 2019
J o m s, 03 Jan 2019Then your GB is just another hair down into the drain. Poor... morelol
There is basically no workload that can 100% parallelize all cores. None.
The closest thing is encoding but with way fewer cores like 3 or 4, not 8.
Try any other cpu benchmark to see it for yourself.
- D
- AnonD-558092
- 45}
- 04 Jan 2019
Anonymous, 03 Jan 2019Because Qualcomm and Sony are pretty close friends. Son... moreSamsung also worked on it. They cross-licence their technologies.
- ?
- Anonymous
- Js{
- 03 Jan 2019
Nirvan2301, 03 Jan 2019I believe that the priced that will tag onto this phone wil... moreXperia X has november patch.
Does A9 2016 have?
The low end L1 has november patch.
Sony has 7 devices with android 9.
- ?
- Anonymous
- Js{
- 03 Jan 2019
AnonD-558092, 03 Jan 2019Why people congratulate Sony? As far as I know, this chip w... moreBecause Qualcomm and Sony are pretty close friends.
Sony helped Qualcomm to develop the 845. Mainly the Spectra.
Sony has been part of the development of 855 as well.
And projects related to 5G...
Why do you think Cristiano Amon is always at Sony's MWC press?