“Very solid” as in catching up with the market leader’s model from 2 generations ago? Would AMD have a “very solid” CPU if it was outperformed by the same kind of margin by Intel’s CPUs? Or same but for AMD/Nvidia?
But that's because IPC improvements since Sandy Bridge have been fairly minimal, which is why it's a huge very solid improvement. Had every iteration since Sandy Bridge involved Apple-like improvements going from the A5 back in 2011 to the A12 today, then yes, catching up to 5 year old IPC would be pretty unimpressive.
You'll have to look at SUSTAINED performance, especially multicore performance. In that department, ARM was never too far behind. This is mobile device. It doesn't really matter what you can do with 3.6W single core when 5W is the realistic thermal envelope.
Moreover i wonder how these power measures are done. There isn't much in this article about this. I really doubt these data show real core power consumption, there are a lot of power pins on the SOC and a reliable measure of the core wattage is nearly impossible. I think these values must be increased a bit.....or a lot more.
Same happen on desktop CPUs. Serious testers adopt the "at the wall" power parameter like the only that give a real figure of the "system" power consumption under cpu test torture. Measure the core power draw is really hard or impossible outside factories.
Ting Cao et al (http://users.cecs.anu.edu.au/~steveb/pubs/papers/y... and other related papers) use a Hall Effect sensor between the power supply and voltage regulator for the chip. Not sure whether this methodology can be used for these SOCs.
Not really. It’s also not quite equal to an A series three generations back - it’s still a bit behind that. So I imagine that Apple will have better performance next year, as usual, and these chips will be about as far behind.
It’s only when Apple’s chips are taken out of the equation can this be called good. Otherwise, it’s terrible. That’s called the testing on a curve everyone else gets. Take the best out, and everyone gets to look good, even though they aren’t.
Ryzen IPC indeed very solid. Ryzen 1xxx IPC is between Skylake and Kabylake. Only its clock headroom was its limitation. The upcoming Ryzen 3xxx series built upon 7nm FinFET is rumored to beat any production Intel CPU in IPC.
I thought Zen/Zen+ (Ryzen 1xxx/2xxx) IPC was closer to Haswell/Broadwell. Also, I don't recall any significant IPC improvement between Skylake and Kaby Lake. Weren't the improvements centered around frequency and power efficiency? Also, I'm going to request a source on that Zen2 (Ryzen 3xxx) IPC improvement. We do know that they've improved the AVX engines pretty significantly, but that will only help specific workloads and it is still a narrower implementation than Intel's current AVX engines. IPC improvements beyond that haven't really been quantified as far as I'm aware.
The 7nm process doesn't inherently affect IPC. It should help with size, frequency, and power efficiency, but there is no guarantee that Zen2 will hit the same frequencies as Intel's mainstream chips. Even at 7nm 5GHz isn't exactly easy without making some tradeoffs AMD may not be willing to make.
Benchmarking is really not a great way to test different platforms against one another. Not an exact science and there are dozens upn dozens of different types. Above you see only two.
I am not suggesting this or any other website change methodology, just that the test results above are far better at judging A10 vs A 11 vs A12 and various ARM CPUs on Android against other ARM CPU's on Android.
Indeed, benchmark scores depend a lot on compilers, OS and system libraries. Geekbench and SPEC spend a significant amount of time in malloc, memcpy, memset, sin, cos, exp, pow and log. The next Android version has much faster math functions which will improve floating point performance.
And you did explain the perf/watt is about equal, so that is now very solid and state of the art. Also thanks for explaining that Geekbench and Spec are comparable to desktop workloads, as they have a greater memory footprint. Not every test is Antutu, and I wish people would stop repeating that claim also.
GB is better than antutu, but it's still too small a load. He appears to be saying that in terms of approximating actual workloads Speedometer > spec > gb > antutu.
"The point I’m trying to make here is that the vast majority of real-world applications behave a lot more like SPEC than GeekBench4: Most notably Apple’s new A12 as well as Samsung’s Exynos 9810 contrast themselves in the two extremes as shown above. In more representative benchmarks such as browser JS framework performance tests (Speedometer 2.0), or on the Android side, PCMark 2.0, we see even greater instruction and data pressure than in SPEC – multiplying the differences exposed by SPECfp."
A single perf/watt figure is not a great way to characterize a CPU. It only applies to a specific implementation of a product, and even then could be changed by a software update or even a settings change. CPUs have a curve of maximum performance available at each power level. And that is never a linear relationship. So saying A12 is 60% faster at 60% more power use is really comparing apples to oranges. If A12 was scaled down to Kirin 980 levels the perf/watt would most likely increase because voltage could be lowered and memory latency costs less wasted CPU cycles.
I think the dual aligned bar charts is a very misleading way to show this information. A simple scatter plot of performance v. power is much better. Even better would be to test each CPU at a range of power levels and plot the curve for each one.
I'd actually be fine with a simple bar chart of 'xx spec score perf / watt', supplemented with a scatterplot to clarify the absolute values involved. But the double-bar chart with multiple units there conveys basically nothing. Please, Andrei, if you're going to talk about efficiency, put up something quantitative about it instead of waving hands at a picture.
I get what it's saying - but I also agree that an efficiency chart of Joules / Performance would help.
One key takeaway that might be unclear is that while Apple's devices do indeed use more power *instantaneously* (Watts), they may end up using almost the same amount of _energy_ as the Kirin, since they finish proportionately faster - with the bonus that you get the result faster, too.
If what you're doing is already plenty fast, that distinction won't matter; but on Javascript-heavy websites you can still notice a slowdown on less-capable processors. Also, time the CPU is in use is quite likely to be time that the screen is on, a factor weighing against less-powerful processors.
At the end of the day, a faster CPU using the same level of energy may mean you can get away with slightly less battery - which might well be a key driver for Apple, since they want thin and light.
That is just silly. Optimizations exist. Its like testing GPU gaming performance on Mac vs. PC. On the surface with the exact same CPU and RAM, PC scores insanely faster than Mac. Not at all the same, not even close.
Not saying you did anything wrong, its just that cross platform results should always be taken with a grain of salt and not as the end all be of of performance.
spec06 test is just source code, you need to compile it and ran native binary code on different platform. So it has just two difference:1.compiler,2.cpu. game is another thing, most gpu benchmark is close source software, you cannot know what it actually is doing.
Also kernel, drivers and system libraries (I assume spec is targeting posix, so, that means it's issuing syscall directly, using a platform natural wrapper or calling into the libc --- those can all affect the results).
Because most people don’t know code, compilers, kernel vs user space, syscalls, binary vs interpreted or any such things beyond maybe having heard of them.
Meanwhile I suppose one could consider the compiler a factor in this - differences between GCC, LLVM and MSVC are sometimes significant, as is potential use of various vector things between platforms, but I expect you keep at least the most obvious and impactful of those in mind when running these tests.
Also want to say: I appreciate your tests and deep knowledge, in no small part because most other reviewers probably never compiled let alone wrote code. Please keep doing this, it is unique and special.
That's not true. Both GB and SPEC do large mallocs which use mmap/sbrk as the underlying implementation, and SPEC benchmarks access various files. If the OS uses a larger page size, TLB misses and pagefaults are significantly reduced, particularly on benchmarks with a large working set like SPEC.
As somebody who has reverse-engineered some of the compute tests in GB, I can assure you those do not use mallocs or *any* syscalls in their timed loops.
That's not true either. GB mallocs the datasets before executing the benchmarks, however several benchmarks use malloc heavily in their inner loops (which means it will make syscalls when it needs more memory). And malloc implementation varies dramatically between Android, iOS and Linux.
Which GB benchmarks are you referring to? Those GB *compute* benchmarks (fp-related) that I've looked at did not use malloc in their timed loops (it would be idiotic if they did).
Malloc is significant in Lua, SQLite, HTML5Dom and LLVM. The latter spends ~20% of total time in malloc. Throughput oriented benchmarks won't use malloc in the inner loop of course, but the statement was about all of SPEC. And the fact is both SPEC and GB make system calls. Note the system calls don't add up to a large amount of time - other OS differences such as pagefault handling and page size make a much larger difference.
I see. So that means most (all?) of GB throughput tests are perfectly fine for benchmarking then? BTW, the sgemm GB test uses a 16KB dataset -- I sincerely doubt OS page handing plays any measurable role in it.
OS pagefault handling matters if you use huge amounts of data and initialize it inside the timing loop - that's the case for SPEC, not GB. However GB is affected by the page size.
Perhaps GB used to do sgemm with 512x512 matrices at the writing of that doc -- they don't now.
Download Geekbench-4.3.1 for linux, md5 58fa9a984992641cd492c14efae895db, the sgemm relevant addresses are at: 0x52d0f0 : sgemm avx 0x52daf0 : sgemm avx+fma
Observe the calls at 0x52d11c and 0x52db1c in the first and the second routine, respectively -- they both call a malloc wrapper with a tad more (for alignment purposes) than 16KB.
You're confused with the traditional blocking optimization which processes the matrices in smaller parts to reduce cachemisses. The source code for an early GB4 shows SGEMM uses 3 matrices of 768x768, so 6.75MB in total. It's possible it was reduced in a later version, or the doc is out of date.
Ok, I finally sat down and traced it (as opposed to looking at objdumps) and you're correct -- their current sgemm matrix width is still 768 columns even in the current version. Which means the 16K must be a transit buffer. They also use a registers-only inner kernel in that multiplication, which is why I originally took the 16K was the entire dataset. My bad.
Bench marking can be very useful and a great way to see how hardware compares to each other. But it is not exact or even close to that. I say this because you can go from review to review and get very different results with the same hardware being tested because of how the reviewer chose to set things up.
Example: CPU testing can have very different results from different websites depending on how the reviewer set everything up. Heck using one board form another can create different results because of how some board makers like to try to cheat the numbers by changing the bus speed slightly or setting certain CPU settings that enhance the performance as in run the CPU out of spec by default with the Auto setting.
So basically a site like Anandtech which can be pretty much trusted to throw out reviews that have solid numbers and the odd time they mess things up they are the first to pull the review or at least correct it which is why I trust their reviews.
You are right, of course, but what they *want* to say is that better benchmarks don't always translate to a better subjective experience, which is what matters most to the end user.
Finnish computer magazine "Mikro Bitti "did test gaming phones and after 30min play time many phones did drop to 60% from normal speed... Some could remain within 20% or normal speed. So long term test are really needed to those who us phone long period at time! Quite many phone were just fine about 10-15 min after that there was really bad drop in speed. We need tests that last much longer than they do normally, to really test phones and SOCs in general!
I agree. While it's great that this is a big leap over previous Android SoCs, and that warrants some applause, it's utterly shocking that no one can catch up to even a 2 year old Apple SoC.... I'm an Android user, so I have no choice but to pick from the list of subpar (relative to Apple) SoCs. I also find it frustrating to see Samsung, Huawei and Qualcomm constantly boast about their SoCs.... only to find that the are still 2 years behind Apple. I wish Nvidia came back. Their X1 was a serious powerhouse for the time.
I don't agree - it's very clear Kirin 980 does catch up with A10. Snapdragon 8150 will be clocked 10% higher so should leave A10 behind, and the laptop 8180 variant is clocked at 3GHz, so should get well above A10 and close to A11 performance.
Also remember that the A12 has 4 times the amount of cache. If one quadruples the L3 in the 980 to 16MB then SPEC and GB4 performance will improve a lot. Although the total core+cache area would still be smaller than A12, nobody is willing to spend that much area/cost.
Go south instead. A12 Big core uses 1.7x(int)/1.6x(FP) power of an much higher clocked A76 in Kirin 980. When you put a ball down tu sustainable limit for a pair of them A76's stay at around 2GHz losing only 23% of benchmark performance while A12's need to downclock in more then a half meaning performance is the same while A76 is much cheaper to produce. When you add into calculation the better scheduler approaches than what Huawei used, more RAM and better memory scheduling on android and also better cooling solution's available from different vendors on Android phones the real world usage will actually be better/faster on Android.
But why do you claim this isn't solid performance? Kirin 980 has better perf/W than A12, far better perf/area, so lower cost, how is that not extremely solid?
It's good - but if your screen and RAM is on for an extra second waiting for that website to fully load, you're wiping out that efficiency game, let alone your life spent waiting for it. (And I'm an Android user, but I can see the benefit.)
>but if your screen and RAM is on for an extra second waiting for that website to fully load, you're wiping out that efficiency game
Well yeah, they have very similar efficiency. That's literally shown right on the SPEC chart. The A12 uses higher average power but takes less time, which works out to basically the same energy.
Subpar SoCs and all these benchmarks. I've been running Android since Froyo 2.2 and until now 7.1.2 all customROMs and custom kernel. I used the first iPhone and yes that time it was something great later after the Snapdragon S4 Pro it was great for Android and after the 820, I honestly can't see any difference in comparing the SoCs where the A series can only Run th jailed iOS where user can't do anything apart from being rules by Orwellian Apple ecosystem. And people are like ashamed to use despite not seeing any masssssive UX loss.
My phone has 820 and OP3T has 821 which uses highest Clock speed on high perf cores which caused instability and guess who brought to this to attention ? Sultanxda and it was implemented by Flar2 and all devs. Guess what ? It lost 200-300MHz clock speed. And judging by this severe fetishistic obsession to benchmarks on smartphones would have caused it to choke and die I suppose ? But guess what ? User experience barely being affected, thanks to custom drivers they build.
And how about LineageOS, Running an OSS tuned for customation and user choice vs the iTunes dependent iDevices and Apple stamp approved.
Its really a shame that how GB is being used all over as if its some sort of HWBOT points with Cinebench or WPrime. On smartphones we need UX and user choice. Look at Huawei EMUI garbage their phones are blocked by VLC due to aggressive background killing or blocking Installing Nova launcher and Bootloader unlock blocked.
No surprise how this Apple taking over the world with A series is emphasized everywhere on AT Forums, Various blogs when their A12X is fabled for killing Laptops and Xbox as quoted by Apple without a proper file manager or mouse KB support.
It would be great if Android reviews forlcuaed more on choice vs these bragging rights. As if the Android SoCs are causing some sort of debilitating disability losing to A series. Pathetic throttling isn't mentioned anywhere on iPhone reviews. Caught red handed deep pockets, enable a switch to user who is ultra dumb to know what it does.
So, has a76 caught up to the a10 (I've forgotten the name of the cores) or not? You seem to be saying both of those ("[...] no one can catch up to even a 2 year [...]" and "[...] only to find that the [sic] are 2 years behind [...]"). If it makes you feel better this is the closest Android has been to Apple since... 2013 (whatever year they went v8a) they are probably quite close to the limits of what standard architectures and our (mass) manufacturing capabilities can achieve. Instead, as had been noted elsewhere, you'll see further work with accelerators & heterogenous systems architecture.
Its kind of a moot point if you can't buy those faster Apples chips to use in anything else, its basically Apple saying "see, our stuffs so fast! No you cant play with it or buy it to use for yourself, and no we aren't going to let you install your stuff on our devices"
How? I have an iPhone, but I´d still _love_ for the competition to catch up. Talking Android SoCs up as "decent" when they´re anything but just tells the manufacturers that they don´t really need to try harder. So Android gets shit SoCs and Apple can price their stuff above $1k; I don´t see any group of customers here as "winners".
I use both Android and iPhone on a daily basis. There is not a single thing I do or app I use where there is any meaningful difference between the platforms in performance apart from my trading app where iOS kills off its background pricing feeds when supposedly multitasking making it next to useless...
Its a moot point in that you're locked into a platform, it means it should only be compared to the other chips within that platform. It doesn't matter how well a more standard ARM core performs next to Apples because you cant use that Apple core in anything that isnt an Apple device and vice versa. So no matter how much you oohh and ahh over which is faster, there will never be a point where its really an issue because any cross platform software will target the lowest common denominator, which will be slower than the fastest chip of either platform. So all this hemming and hawing about "wow no one has caught up with apple" is pointless, because theres never a point where it will be an issue.
Its a moot point in that you're locked into a platform, it means it should only be compared to the other chips within that platform. It doesn't matter how well a more standard ARM core performs next to Apples because you cant use that Apple core in anything that isnt an Apple device and vice versa. So no matter how much you oohh and ahh over which is faster, there will never be a point where its really an issue because any cross platform software will target the lowest common denominator, which will be slower than the fastest chip of either platform. So all this hemming and hawing about "wow no one has caught up with apple" is pointless, because theres never a point where it will be an issue.
ARM has different economic concerns than Apple. They are building compact, reusable cores that can be slapped down in inexpensive (comparatively) SoCs by a multitude of clients. There's no reason to think they couldn't make a CPU that performed similar to an A12, but their business model doesn't lead that direction. They aim for a balance of cost/size, efficiency, and performance when Apple, designing only for premium products, doesn't care if they have a big chip size and comparatively high cost since to fab. The A76 looks very good from ARM's perspective seeing as it'll be used by a variety of vendors delivering SoCs using these cores in products for that can cost 2/3 of an iPhone's asking price.
It catches up with current generation of Apples ARM core's when it comes to the performance/W metric and still being significantly higher clocked. This also means that sustainable performance is the same. The Huawei SoC uses EAV+ scheduler & QC timed window approach actually gives a healthy performance difference in fast switching workloads (15~20%). All together A76 is a better design for mobile.
From a power efficiency standpoint, It doesn't really matter. Both my LG V30 and my iPhone XS Max get 2-3 days on a charge. CPU performance has also stopped mattering for me years ago. I only made the jump to the iPhone to ensure knowledge of both architectures, be able to do iOS development and testing, and have 512 gb of built in storage. Then again, I'm not a gamer, and I don't try to use my phone as a replacement for a desktop or laptop PC.
A76 looks like a good improvement, but it will be limited to phones and tablets. I would love for someone to make something similar to the rasberry pi or beagleboard, a small credit card PC with a snapdragon 8150 or 845 and lots of RAM on boards for fun projects.
There are people who make project boards with high end ARM processors and several GB of ram. You never heard about them being used as a hobby board because they cost as much (or more) as a bargin cellphone with the same chip inside it. Mostly due to low production quantities so they don't get as big of a discount, and also because most hobby projects that would need the processing power of that sort of board would probably also need the additional peripherals that come in a phone as well (display, camera, etc)
I don't think it will be limited to phones and tablets.
This is a huge step forward compared to Cortext-A72 which is used in RK3399 which is used in a lot of home-server oriented SBCs and even laptops (see Pinebook Pro).
No A12X in the benchmarks? I thought AnandTech had done at least a preliminary look into the A12X, but it turns out that it hasn't happened yet. Is there an iPad Pro review in the works?
Brett is working on the iPad Pro review. Performance wise the A12X isn't much different to the A12, there's just two more cores and a bit more bandwidth.
@ Andrei: Thanks for this A76 follow-up from the Kirin review. Question: How much of a role in the memory subsystem bottleneck does the latency and throughput of the DRAM and DRAM bus play, especially in light of the the fruity company's A11 and A12/A12x SoCs performance advantage? I know Apple is notoriously tight-lipped about details, but I recall reading some reports about a 128 bit wide memory bus used in the iPhone X and its successors, which would move data a lot faster. Just wondering how much performance is left on the table by Android devices by their current use of narrower memory interfaces (and cheaper/slower memory). Any information or informed guess is appreciated.
Apple has the same 4x16b interface as all flagships SoCs in mobile, there's no difference in width or the DRAM itself. The microarchitecture of the memory controllers and SoC memory subsystem obviously is going to be very different.
Hi Andrei, great article. Quick question, can you clarify what was the other edge you had in mind when you said "[...]for the Snapdragon 845 this was a double-edged sword as memory latency saw a degradation over the Snapdragon 835. This degradation seemingly caused the Cortex A75 in the S845 to maybe not achieve its full potential"
Maybe streaming loads improved? They must've simulated this beforehand so either something did improve or there was an implementation issue.
The goal of a system cache is to improve power and performance by lowering latency. Every extra level of cache needs to be 4-8 times larger than the previous level or it won't work well (eg. Kirin 980 64+64KB L1, 512KB L2, 4MB L3). A system cache likely improves power, but it seemed to add a lot of latency and thus performance was worse on the 845. Hence double-edged sword.
Hi Andrei, great article. Quick question, can you clarify what was the other edge you had in mind when you said "[...]for the Snapdragon 845 this was a double-edged sword as memory latency saw a degradation over the Snapdragon 835. This degradation seemingly caused the Cortex A75 in the S845 to maybe not achieve its full potential"
Maybe streaming loads improved? They must've simulated this beforehand so either something did improve or there was an implementation issue.
Since I do not play games on my phone, for the last 4 years I have been buying $200 mid range phones. But I am tired of the same A53-based SoCs over and over for so many years! Would really like to see what difference A56 vs. A53 brings to the table. Yeah, A76 is the new cool thing, but we all know that the A53 sold much much more than A73, if the A56 is what is going to be featured in our TV boxes and sticks and SBCs for the next 5 years or so, would be good to know what to expect from it!
Thanks for the write-up Andrei, would be interested in a further exploration of how the ARM A76 core compares with the Intel/AMD desktop architectures. GB scores for the mate 20 pro put it at ~Ryzen 2700U levels but obviously thats far from the whole story...
Please educate me on the "idiocy" of taking an in-depth look at how A76 compares with Intel/AMD cores in Front End/Back End/Memory subsystems with indepth comparisons (I. E not geekbench scores)? Given the increasing adoption of ARM in higher-TDP usecases it'd be fairly enlightening to see how the best of ARM ex-apple holds up to the leading x86 architectures in a variety of tasks.
@Andrei Hmm, the A12 consumed 10% more Joules but had 60% more perf ... I think ARM has a ways to go here. Power usage is proportional but A12 is faster at completing the task, significantly. Normalizing for energy used as opposed to watts would be more informative, as batteries are measured in W*h = Joules not Watts.
Actually in performance/W (in the table performance/J) it's other way around even A76 is significantly higher clocked. If you put both in the sustainable performance metrics pair of A76 will perform the same as A12 two big ones. If you put them both to sustainable optimal leakage clocks for FinFET structures (1.6~1.7GHz) A76 will end using less than a half power of an A12 big core. Apple big cores in their current state are in advantage to the ARM's A76 but not on the mobile phones allowed DTP.
Burst performance is what matters here though: how often are you doing long sustained computation on a phone other than gaming? In that sense, 2x W for > 2x perf is a totally acceptable tradeoff to get perf/J up. I think A76 is optimizing for the wrong thing and it shows here.
A76 for midrange devices is what I'm looking forward to. A cheap 670 successor with 4x A55 and 2x A76 with a total device price of $200 sounds good. I just need something to replace my ancient SD650 which surprisingly still rocks after almost three years.
The S675 is already announced. Dream on about $200 phone with it, it ain't gonna happen as their is no unhealthily competition like it whose back in the days between Xiaomi & LeEco. At 250~280$ we may see cuple models.
Doesn't seem much of a jump, next thing after A76 will be a true next gen.
But hey, smartphones are already fast enough for every mobile use.
My S9 can do:
1080p60fps youtube 1080p h264/HEVC anime bd rips As fast as it can get for browsing
Smartphones reached the point of being fast enough, sales will suffer a decline in the upcoming years unless people are stupid enough to keep being influenced by "you need a new flagship every 18 months because it's better and you need it"
@Andrei Frumusanu Another excellent analysis, thank you. Q: If you could have your own benchmark suite (and not just what is available now) , what would you test and how on current/future mobile SoCs?
I would have loved to see a comparison (however flawed) in performance and maybe energy efficiency against a low end laptop processor - a Celeron maybe. I loved the A12 performance comparison versus Intel cores (not full processors). I find it relevant considering that ARM is pushing for always-on, always-connected computers. Also, how about an article about why Apple is (or isn't) working for A12-based MacBook Air and so on?
I read the Mate 20 Pro review, but I appreciate this article as the (very interesting!) CPU testing drowned a bit in the massive camera testing portion.
I'm wondering, though, will you be testing the A12X/iPad Pro any time soon? Would be extremely interesting to see this same type of testing run there, in particular regarding power draw and efficiency.
$200 may be a tad optimistic. $250 phones now use SD636, $300 gets you SD660 and $350+ gets you SD710. That's ridiculous product segmentation in my eyes but Qualcomm knows best... $250 for SD67x and $350+ for SD7xx makes more sense.
IOS apps are either written in Swift or Objective-C while the Android space is Java based.
I'm wondering how much the difference in languages matters when it comes to these benchmarks. Back in the day, it mattered quite a bit which language ran on a machine as some compilers were better suited to the task than others. Unless the benchmarks were written by equally competent assembly language coders (unlikely) it's not clear where the Qualcomm and Apple chips stand .
The first slide promises "laptop-class performance" for the A76. That intrigued me. All the ARM option laptops I've seen were underpowered up to now and much too costly. ARM laptops seem to compete in the high end for everything (battery life, display, build quality) but often sacrifice storage and performance. I don't think I'll ever use an ARM laptop for the next 10 years. On the other hand, this has me interested in a comparison to the Jaguar cores still used in the modern Sony and MS consoles. They might very well be less powerfull than a Cortex A76, generally speaking. What sort of upgrade path does the Switch have? Nvidia isn't doing any new SOCs, right? I doubt they'll drastically switch (hah) architectures, so something like this might do it.
Any way to make a generational article similar to this? From the first phones, first dual core phones all the way here? Probably pretty difficult with all the different OS revisions and finding working hardware.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
99 Comments
Back to Article
xype - Tuesday, November 20, 2018 - link
“Very solid” as in catching up with the market leader’s model from 2 generations ago? Would AMD have a “very solid” CPU if it was outperformed by the same kind of margin by Intel’s CPUs? Or same but for AMD/Nvidia?TheinsanegamerN - Tuesday, November 20, 2018 - link
Ryzen was catching up to 5 year old intel IPC, and was described as "very solid" indeed. So, yeah, it would be.OddFriendship8989 - Tuesday, November 20, 2018 - link
But that's because IPC improvements since Sandy Bridge have been fairly minimal, which is why it's a huge very solid improvement. Had every iteration since Sandy Bridge involved Apple-like improvements going from the A5 back in 2011 to the A12 today, then yes, catching up to 5 year old IPC would be pretty unimpressive.levizx - Wednesday, November 21, 2018 - link
You'll have to look at SUSTAINED performance, especially multicore performance. In that department, ARM was never too far behind. This is mobile device. It doesn't really matter what you can do with 3.6W single core when 5W is the realistic thermal envelope.Samus - Wednesday, November 21, 2018 - link
That is just not true, so you will have to “look it up” for us.Gondalf - Wednesday, November 21, 2018 - link
Moreover i wonder how these power measures are done. There isn't much in this article about this. I really doubt these data show real core power consumption, there are a lot of power pins on the SOC and a reliable measure of the core wattage is nearly impossible.I think these values must be increased a bit.....or a lot more.
Same happen on desktop CPUs. Serious testers adopt the "at the wall" power parameter like the only that give a real figure of the "system" power consumption under cpu test torture. Measure the core power draw is really hard or impossible outside factories.
rgarner - Wednesday, December 12, 2018 - link
Ting Cao et al (http://users.cecs.anu.edu.au/~steveb/pubs/papers/y... and other related papers) use a Hall Effect sensor between the power supply and voltage regulator for the chip. Not sure whether this methodology can be used for these SOCs.melgross - Thursday, November 22, 2018 - link
Not really. It’s also not quite equal to an A series three generations back - it’s still a bit behind that. So I imagine that Apple will have better performance next year, as usual, and these chips will be about as far behind.It’s only when Apple’s chips are taken out of the equation can this be called good. Otherwise, it’s terrible. That’s called the testing on a curve everyone else gets. Take the best out, and everyone gets to look good, even though they aren’t.
kuttan - Sunday, November 25, 2018 - link
Ryzen IPC indeed very solid. Ryzen 1xxx IPC is between Skylake and Kabylake. Only its clock headroom was its limitation. The upcoming Ryzen 3xxx series built upon 7nm FinFET is rumored to beat any production Intel CPU in IPC.BurntMyBacon - Wednesday, November 28, 2018 - link
I thought Zen/Zen+ (Ryzen 1xxx/2xxx) IPC was closer to Haswell/Broadwell.Also, I don't recall any significant IPC improvement between Skylake and Kaby Lake. Weren't the improvements centered around frequency and power efficiency?
Also, I'm going to request a source on that Zen2 (Ryzen 3xxx) IPC improvement. We do know that they've improved the AVX engines pretty significantly, but that will only help specific workloads and it is still a narrower implementation than Intel's current AVX engines. IPC improvements beyond that haven't really been quantified as far as I'm aware.
The 7nm process doesn't inherently affect IPC. It should help with size, frequency, and power efficiency, but there is no guarantee that Zen2 will hit the same frequencies as Intel's mainstream chips. Even at 7nm 5GHz isn't exactly easy without making some tradeoffs AMD may not be willing to make.
goatfajitas - Tuesday, November 20, 2018 - link
Benchmarking is really not a great way to test different platforms against one another. Not an exact science and there are dozens upn dozens of different types. Above you see only two.blu42 - Tuesday, November 20, 2018 - link
What would you suggest Andrei ran instead of benchmarks?goatfajitas - Tuesday, November 20, 2018 - link
I am not suggesting this or any other website change methodology, just that the test results above are far better at judging A10 vs A 11 vs A12 and various ARM CPUs on Android against other ARM CPU's on Android.Wilco1 - Tuesday, November 20, 2018 - link
Indeed, benchmark scores depend a lot on compilers, OS and system libraries. Geekbench and SPEC spend a significant amount of time in malloc, memcpy, memset, sin, cos, exp, pow and log. The next Android version has much faster math functions which will improve floating point performance.melgross - Thursday, November 22, 2018 - link
When the devices that are benchmarked are tested with real apps, the results come out about the same. What more do you expect?Andrei Frumusanu - Tuesday, November 20, 2018 - link
> Benchmarking is really not a great way to test different platforms against one another.> Not an exact science
This is just a stupid claim. Benchmarking *is* exact and there is nothing wrong with cross platform tests.
Alistair - Tuesday, November 20, 2018 - link
And you did explain the perf/watt is about equal, so that is now very solid and state of the art. Also thanks for explaining that Geekbench and Spec are comparable to desktop workloads, as they have a greater memory footprint. Not every test is Antutu, and I wish people would stop repeating that claim also.tuxRoller - Tuesday, November 20, 2018 - link
GB is better than antutu, but it's still too small a load.He appears to be saying that in terms of approximating actual workloads Speedometer > spec > gb > antutu.
"The point I’m trying to make here is that the vast majority of real-world applications behave a lot more like SPEC than GeekBench4: Most notably Apple’s new A12 as well as Samsung’s Exynos 9810 contrast themselves in the two extremes as shown above. In more representative benchmarks such as browser JS framework performance tests (Speedometer 2.0), or on the Android side, PCMark 2.0, we see even greater instruction and data pressure than in SPEC – multiplying the differences exposed by SPECfp."
stepz - Wednesday, November 21, 2018 - link
A single perf/watt figure is not a great way to characterize a CPU. It only applies to a specific implementation of a product, and even then could be changed by a software update or even a settings change. CPUs have a curve of maximum performance available at each power level. And that is never a linear relationship. So saying A12 is 60% faster at 60% more power use is really comparing apples to oranges. If A12 was scaled down to Kirin 980 levels the perf/watt would most likely increase because voltage could be lowered and memory latency costs less wasted CPU cycles.I think the dual aligned bar charts is a very misleading way to show this information. A simple scatter plot of performance v. power is much better. Even better would be to test each CPU at a range of power levels and plot the curve for each one.
ABR - Wednesday, November 21, 2018 - link
I'd actually be fine with a simple bar chart of 'xx spec score perf / watt', supplemented with a scatterplot to clarify the absolute values involved. But the double-bar chart with multiple units there conveys basically nothing. Please, Andrei, if you're going to talk about efficiency, put up something quantitative about it instead of waving hands at a picture.GreenReaper - Wednesday, November 21, 2018 - link
I get what it's saying - but I also agree that an efficiency chart of Joules / Performance would help.One key takeaway that might be unclear is that while Apple's devices do indeed use more power *instantaneously* (Watts), they may end up using almost the same amount of _energy_ as the Kirin, since they finish proportionately faster - with the bonus that you get the result faster, too.
If what you're doing is already plenty fast, that distinction won't matter; but on Javascript-heavy websites you can still notice a slowdown on less-capable processors. Also, time the CPU is in use is quite likely to be time that the screen is on, a factor weighing against less-powerful processors.
At the end of the day, a faster CPU using the same level of energy may mean you can get away with slightly less battery - which might well be a key driver for Apple, since they want thin and light.
goatfajitas - Tuesday, November 20, 2018 - link
That is just silly. Optimizations exist. Its like testing GPU gaming performance on Mac vs. PC. On the surface with the exact same CPU and RAM, PC scores insanely faster than Mac. Not at all the same, not even close.Not saying you did anything wrong, its just that cross platform results should always be taken with a grain of salt and not as the end all be of of performance.
deathBOB - Tuesday, November 20, 2018 - link
Games use abstractions like DirectX. It’s not a good comparison.ternnence - Tuesday, November 20, 2018 - link
spec06 test is just source code, you need to compile it and ran native binary code on different platform. So it has just two difference:1.compiler,2.cpu. game is another thing, most gpu benchmark is close source software, you cannot know what it actually is doing.tuxRoller - Tuesday, November 20, 2018 - link
Also kernel, drivers and system libraries (I assume spec is targeting posix, so, that means it's issuing syscall directly, using a platform natural wrapper or calling into the libc --- those can all affect the results).Andrei Frumusanu - Tuesday, November 20, 2018 - link
The suite is void of any system calls during the main runtime. I don't understand why people keep arguing about non issues.jospoortvliet - Tuesday, November 20, 2018 - link
Because most people don’t know code, compilers, kernel vs user space, syscalls, binary vs interpreted or any such things beyond maybe having heard of them.Meanwhile I suppose one could consider the compiler a factor in this - differences between GCC, LLVM and MSVC are sometimes significant, as is potential use of various vector things between platforms, but I expect you keep at least the most obvious and impactful of those in mind when running these tests.
Also want to say: I appreciate your tests and deep knowledge, in no small part because most other reviewers probably never compiled let alone wrote code. Please keep doing this, it is unique and special.
Wilco1 - Tuesday, November 20, 2018 - link
That's not true. Both GB and SPEC do large mallocs which use mmap/sbrk as the underlying implementation, and SPEC benchmarks access various files. If the OS uses a larger page size, TLB misses and pagefaults are significantly reduced, particularly on benchmarks with a large working set like SPEC.blu42 - Wednesday, November 21, 2018 - link
As somebody who has reverse-engineered some of the compute tests in GB, I can assure you those do not use mallocs or *any* syscalls in their timed loops.Wilco1 - Wednesday, November 21, 2018 - link
That's not true either. GB mallocs the datasets before executing the benchmarks, however several benchmarks use malloc heavily in their inner loops (which means it will make syscalls when it needs more memory). And malloc implementation varies dramatically between Android, iOS and Linux.blu42 - Wednesday, November 21, 2018 - link
Which GB benchmarks are you referring to? Those GB *compute* benchmarks (fp-related) that I've looked at did not use malloc in their timed loops (it would be idiotic if they did).Wilco1 - Wednesday, November 21, 2018 - link
Malloc is significant in Lua, SQLite, HTML5Dom and LLVM. The latter spends ~20% of total time in malloc. Throughput oriented benchmarks won't use malloc in the inner loop of course, but the statement was about all of SPEC. And the fact is both SPEC and GB make system calls. Note the system calls don't add up to a large amount of time - other OS differences such as pagefault handling and page size make a much larger difference.blu42 - Thursday, November 22, 2018 - link
I see. So that means most (all?) of GB throughput tests are perfectly fine for benchmarking then? BTW, the sgemm GB test uses a 16KB dataset -- I sincerely doubt OS page handing plays any measurable role in it.Wilco1 - Thursday, November 22, 2018 - link
Nothing in GB uses a tiny 16KB dataset - workload memory usage is around 32-64MBytes. A single SGEMM matrix is 1MB: https://www.geekbench.com/doc/geekbench4-cpu-workl...OS pagefault handling matters if you use huge amounts of data and initialize it inside the timing loop - that's the case for SPEC, not GB. However GB is affected by the page size.
blu42 - Thursday, November 22, 2018 - link
Perhaps GB used to do sgemm with 512x512 matrices at the writing of that doc -- they don't now.Download Geekbench-4.3.1 for linux, md5 58fa9a984992641cd492c14efae895db, the sgemm relevant addresses are at:
0x52d0f0 : sgemm avx
0x52daf0 : sgemm avx+fma
Observe the calls at 0x52d11c and 0x52db1c in the first and the second routine, respectively -- they both call a malloc wrapper with a tad more (for alignment purposes) than 16KB.
Wilco1 - Thursday, November 22, 2018 - link
You're confused with the traditional blocking optimization which processes the matrices in smaller parts to reduce cachemisses. The source code for an early GB4 shows SGEMM uses 3 matrices of 768x768, so 6.75MB in total. It's possible it was reduced in a later version, or the doc is out of date.blu42 - Thursday, November 22, 2018 - link
Ok, I finally sat down and traced it (as opposed to looking at objdumps) and you're correct -- their current sgemm matrix width is still 768 columns even in the current version. Which means the 16K must be a transit buffer. They also use a registers-only inner kernel in that multiplication, which is why I originally took the 16K was the entire dataset. My bad.Wilco1 - Thursday, November 22, 2018 - link
So it's the document that was out of date then. Thanks for double checking.tuxRoller - Wednesday, November 21, 2018 - link
How are the processes created?rocky12345 - Tuesday, November 20, 2018 - link
Bench marking can be very useful and a great way to see how hardware compares to each other. But it is not exact or even close to that. I say this because you can go from review to review and get very different results with the same hardware being tested because of how the reviewer chose to set things up.Example: CPU testing can have very different results from different websites depending on how the reviewer set everything up. Heck using one board form another can create different results because of how some board makers like to try to cheat the numbers by changing the bus speed slightly or setting certain CPU settings that enhance the performance as in run the CPU out of spec by default with the Auto setting.
So basically a site like Anandtech which can be pretty much trusted to throw out reviews that have solid numbers and the odd time they mess things up they are the first to pull the review or at least correct it which is why I trust their reviews.
Amandtec - Wednesday, November 21, 2018 - link
You are right, of course, but what they *want* to say is that better benchmarks don't always translate to a better subjective experience, which is what matters most to the end user.Ironchef3500 - Wednesday, November 21, 2018 - link
:)haukionkannel - Thursday, November 22, 2018 - link
Finnish computer magazine "Mikro Bitti "did test gaming phones and after 30min play time many phones did drop to 60% from normal speed... Some could remain within 20% or normal speed. So long term test are really needed to those who us phone long period at time!Quite many phone were just fine about 10-15 min after that there was really bad drop in speed.
We need tests that last much longer than they do normally, to really test phones and SOCs in general!
syxbit - Tuesday, November 20, 2018 - link
I agree. While it's great that this is a big leap over previous Android SoCs, and that warrants some applause, it's utterly shocking that no one can catch up to even a 2 year old Apple SoC....I'm an Android user, so I have no choice but to pick from the list of subpar (relative to Apple) SoCs.
I also find it frustrating to see Samsung, Huawei and Qualcomm constantly boast about their SoCs.... only to find that the are still 2 years behind Apple.
I wish Nvidia came back. Their X1 was a serious powerhouse for the time.
Wilco1 - Tuesday, November 20, 2018 - link
I don't agree - it's very clear Kirin 980 does catch up with A10. Snapdragon 8150 will be clocked 10% higher so should leave A10 behind, and the laptop 8180 variant is clocked at 3GHz, so should get well above A10 and close to A11 performance.Also remember that the A12 has 4 times the amount of cache. If one quadruples the L3 in the 980 to 16MB then SPEC and GB4 performance will improve a lot. Although the total core+cache area would still be smaller than A12, nobody is willing to spend that much area/cost.
ZolaIII - Tuesday, November 20, 2018 - link
Go south instead. A12 Big core uses 1.7x(int)/1.6x(FP) power of an much higher clocked A76 in Kirin 980. When you put a ball down tu sustainable limit for a pair of them A76's stay at around 2GHz losing only 23% of benchmark performance while A12's need to downclock in more then a half meaning performance is the same while A76 is much cheaper to produce. When you add into calculation the better scheduler approaches than what Huawei used, more RAM and better memory scheduling on android and also better cooling solution's available from different vendors on Android phones the real world usage will actually be better/faster on Android.xype - Wednesday, November 21, 2018 - link
"Kirin 980 does catch up with A10", "laptop 8180 variant ", "close to A11 performance"Thanks for agreeing with me.
Wilco1 - Wednesday, November 21, 2018 - link
But why do you claim this isn't solid performance? Kirin 980 has better perf/W than A12, far better perf/area, so lower cost, how is that not extremely solid?GreenReaper - Wednesday, November 21, 2018 - link
It's good - but if your screen and RAM is on for an extra second waiting for that website to fully load, you're wiping out that efficiency game, let alone your life spent waiting for it. (And I'm an Android user, but I can see the benefit.)skavi - Wednesday, November 21, 2018 - link
>but if your screen and RAM is on for an extra second waiting for that website to fully load, you're wiping out that efficiency gameWell yeah, they have very similar efficiency. That's literally shown right on the SPEC chart. The A12 uses higher average power but takes less time, which works out to basically the same energy.
Quantumz0d - Tuesday, November 20, 2018 - link
Subpar SoCs and all these benchmarks. I've been running Android since Froyo 2.2 and until now 7.1.2 all customROMs and custom kernel. I used the first iPhone and yes that time it was something great later after the Snapdragon S4 Pro it was great for Android and after the 820, I honestly can't see any difference in comparing the SoCs where the A series can only Run th jailed iOS where user can't do anything apart from being rules by Orwellian Apple ecosystem. And people are like ashamed to use despite not seeing any masssssive UX loss.My phone has 820 and OP3T has 821 which uses highest Clock speed on high perf cores which caused instability and guess who brought to this to attention ? Sultanxda and it was implemented by Flar2 and all devs. Guess what ? It lost 200-300MHz clock speed. And judging by this severe fetishistic obsession to benchmarks on smartphones would have caused it to choke and die I suppose ? But guess what ? User experience barely being affected, thanks to custom drivers they build.
And how about LineageOS, Running an OSS tuned for customation and user choice vs the iTunes dependent iDevices and Apple stamp approved.
Its really a shame that how GB is being used all over as if its some sort of HWBOT points with Cinebench or WPrime. On smartphones we need UX and user choice. Look at Huawei EMUI garbage their phones are blocked by VLC due to aggressive background killing or blocking Installing Nova launcher and Bootloader unlock blocked.
No surprise how this Apple taking over the world with A series is emphasized everywhere on AT Forums, Various blogs when their A12X is fabled for killing Laptops and Xbox as quoted by Apple without a proper file manager or mouse KB support.
It would be great if Android reviews forlcuaed more on choice vs these bragging rights. As if the Android SoCs are causing some sort of debilitating disability losing to A series. Pathetic throttling isn't mentioned anywhere on iPhone reviews. Caught red handed deep pockets, enable a switch to user who is ultra dumb to know what it does.
tuxRoller - Tuesday, November 20, 2018 - link
So, has a76 caught up to the a10 (I've forgotten the name of the cores) or not? You seem to be saying both of those ("[...] no one can catch up to even a 2 year [...]" and "[...] only to find that the [sic] are 2 years behind [...]").If it makes you feel better this is the closest Android has been to Apple since... 2013 (whatever year they went v8a) they are probably quite close to the limits of what standard architectures and our (mass) manufacturing capabilities can achieve. Instead, as had been noted elsewhere, you'll see further work with accelerators & heterogenous systems architecture.
artk2219 - Tuesday, November 20, 2018 - link
Its kind of a moot point if you can't buy those faster Apples chips to use in anything else, its basically Apple saying "see, our stuffs so fast! No you cant play with it or buy it to use for yourself, and no we aren't going to let you install your stuff on our devices"xype - Wednesday, November 21, 2018 - link
"Its kind of a moot point "How? I have an iPhone, but I´d still _love_ for the competition to catch up. Talking Android SoCs up as "decent" when they´re anything but just tells the manufacturers that they don´t really need to try harder. So Android gets shit SoCs and Apple can price their stuff above $1k; I don´t see any group of customers here as "winners".
Speedfriend - Wednesday, November 21, 2018 - link
I use both Android and iPhone on a daily basis. There is not a single thing I do or app I use where there is any meaningful difference between the platforms in performance apart from my trading app where iOS kills off its background pricing feeds when supposedly multitasking making it next to useless...artk2219 - Tuesday, November 27, 2018 - link
Its a moot point in that you're locked into a platform, it means it should only be compared to the other chips within that platform. It doesn't matter how well a more standard ARM core performs next to Apples because you cant use that Apple core in anything that isnt an Apple device and vice versa. So no matter how much you oohh and ahh over which is faster, there will never be a point where its really an issue because any cross platform software will target the lowest common denominator, which will be slower than the fastest chip of either platform. So all this hemming and hawing about "wow no one has caught up with apple" is pointless, because theres never a point where it will be an issue.artk2219 - Tuesday, November 27, 2018 - link
Its a moot point in that you're locked into a platform, it means it should only be compared to the other chips within that platform. It doesn't matter how well a more standard ARM core performs next to Apples because you cant use that Apple core in anything that isnt an Apple device and vice versa. So no matter how much you oohh and ahh over which is faster, there will never be a point where its really an issue because any cross platform software will target the lowest common denominator, which will be slower than the fastest chip of either platform. So all this hemming and hawing about "wow no one has caught up with apple" is pointless, because theres never a point where it will be an issue.gijames1225 - Tuesday, November 20, 2018 - link
ARM has different economic concerns than Apple. They are building compact, reusable cores that can be slapped down in inexpensive (comparatively) SoCs by a multitude of clients. There's no reason to think they couldn't make a CPU that performed similar to an A12, but their business model doesn't lead that direction. They aim for a balance of cost/size, efficiency, and performance when Apple, designing only for premium products, doesn't care if they have a big chip size and comparatively high cost since to fab. The A76 looks very good from ARM's perspective seeing as it'll be used by a variety of vendors delivering SoCs using these cores in products for that can cost 2/3 of an iPhone's asking price.xype - Wednesday, November 21, 2018 - link
If only companies like Samsung or Qualcomm were big enough to afford to create their own, decent ARM implementations. Alas.ZolaIII - Tuesday, November 20, 2018 - link
It catches up with current generation of Apples ARM core's when it comes to the performance/W metric and still being significantly higher clocked. This also means that sustainable performance is the same. The Huawei SoC uses EAV+ scheduler & QC timed window approach actually gives a healthy performance difference in fast switching workloads (15~20%). All together A76 is a better design for mobile.eek2121 - Wednesday, November 21, 2018 - link
From a power efficiency standpoint, It doesn't really matter. Both my LG V30 and my iPhone XS Max get 2-3 days on a charge. CPU performance has also stopped mattering for me years ago. I only made the jump to the iPhone to ensure knowledge of both architectures, be able to do iOS development and testing, and have 512 gb of built in storage. Then again, I'm not a gamer, and I don't try to use my phone as a replacement for a desktop or laptop PC.eek2121 - Wednesday, November 21, 2018 - link
* Not a gamer on phones. Wish there was an edit button.TheinsanegamerN - Tuesday, November 20, 2018 - link
A76 looks like a good improvement, but it will be limited to phones and tablets. I would love for someone to make something similar to the rasberry pi or beagleboard, a small credit card PC with a snapdragon 8150 or 845 and lots of RAM on boards for fun projects.QChronoD - Tuesday, November 20, 2018 - link
There are people who make project boards with high end ARM processors and several GB of ram. You never heard about them being used as a hobby board because they cost as much (or more) as a bargin cellphone with the same chip inside it. Mostly due to low production quantities so they don't get as big of a discount, and also because most hobby projects that would need the processing power of that sort of board would probably also need the additional peripherals that come in a phone as well (display, camera, etc)asoltesz - Saturday, February 23, 2019 - link
I don't think it will be limited to phones and tablets.This is a huge step forward compared to Cortext-A72 which is used in RK3399 which is used in a lot of home-server oriented SBCs and even laptops (see Pinebook Pro).
The_Assimilator - Tuesday, November 20, 2018 - link
Everyone knows that Geekbench is useless trash; why not take a stand by not including it in your benchmark suite going forward?Andrei Frumusanu - Tuesday, November 20, 2018 - link
We don't have it in our benchmark suite...iwod - Wednesday, November 21, 2018 - link
It is not, not since version 4.coder543 - Tuesday, November 20, 2018 - link
No A12X in the benchmarks? I thought AnandTech had done at least a preliminary look into the A12X, but it turns out that it hasn't happened yet. Is there an iPad Pro review in the works?Andrei Frumusanu - Tuesday, November 20, 2018 - link
Brett is working on the iPad Pro review. Performance wise the A12X isn't much different to the A12, there's just two more cores and a bit more bandwidth.eastcoast_pete - Tuesday, November 20, 2018 - link
@ Andrei: Thanks for this A76 follow-up from the Kirin review. Question: How much of a role in the memory subsystem bottleneck does the latency and throughput of the DRAM and DRAM bus play, especially in light of the the fruity company's A11 and A12/A12x SoCs performance advantage? I know Apple is notoriously tight-lipped about details, but I recall reading some reports about a 128 bit wide memory bus used in the iPhone X and its successors, which would move data a lot faster. Just wondering how much performance is left on the table by Android devices by their current use of narrower memory interfaces (and cheaper/slower memory). Any information or informed guess is appreciated.Andrei Frumusanu - Tuesday, November 20, 2018 - link
Apple has the same 4x16b interface as all flagships SoCs in mobile, there's no difference in width or the DRAM itself. The microarchitecture of the memory controllers and SoC memory subsystem obviously is going to be very different.eastcoast_pete - Tuesday, November 20, 2018 - link
Thanks Andrei! Makes sense.tuxRoller - Tuesday, November 20, 2018 - link
Hi Andrei, great article.Quick question, can you clarify what was the other edge you had in mind when you said "[...]for the Snapdragon 845 this was a double-edged sword as memory latency saw a degradation over the Snapdragon 835. This degradation seemingly caused the Cortex A75 in the S845 to maybe not achieve its full potential"
Maybe streaming loads improved? They must've simulated this beforehand so either something did improve or there was an implementation issue.
Wilco1 - Tuesday, November 20, 2018 - link
The goal of a system cache is to improve power and performance by lowering latency. Every extra level of cache needs to be 4-8 times larger than the previous level or it won't work well (eg. Kirin 980 64+64KB L1, 512KB L2, 4MB L3). A system cache likely improves power, but it seemed to add a lot of latency and thus performance was worse on the 845. Hence double-edged sword.tuxRoller - Wednesday, November 21, 2018 - link
Much appreciated!tuxRoller - Tuesday, November 20, 2018 - link
Hi Andrei, great article.Quick question, can you clarify what was the other edge you had in mind when you said "[...]for the Snapdragon 845 this was a double-edged sword as memory latency saw a degradation over the Snapdragon 835. This degradation seemingly caused the Cortex A75 in the S845 to maybe not achieve its full potential"
Maybe streaming loads improved? They must've simulated this beforehand so either something did improve or there was an implementation issue.
mpbello - Tuesday, November 20, 2018 - link
Since I do not play games on my phone, for the last 4 years I have been buying $200 mid range phones.But I am tired of the same A53-based SoCs over and over for so many years! Would really like to see what difference A56 vs. A53 brings to the table. Yeah, A76 is the new cool thing, but we all know that the A53 sold much much more than A73, if the A56 is what is going to be featured in our TV boxes and sticks and SBCs for the next 5 years or so, would be good to know what to expect from it!
unclevagz - Tuesday, November 20, 2018 - link
Thanks for the write-up Andrei, would be interested in a further exploration of how the ARM A76 core compares with the Intel/AMD desktop architectures. GB scores for the mate 20 pro put it at ~Ryzen 2700U levels but obviously thats far from the whole story...Lolimaster - Wednesday, November 21, 2018 - link
Just don't do it, it's idiotic.Most of the load of GB barely touches the tons of hardware accelerated tasks that x86 has built over the years.
unclevagz - Wednesday, November 21, 2018 - link
Please educate me on the "idiocy" of taking an in-depth look at how A76 compares with Intel/AMD cores in Front End/Back End/Memory subsystems with indepth comparisons (I. E not geekbench scores)? Given the increasing adoption of ARM in higher-TDP usecases it'd be fairly enlightening to see how the best of ARM ex-apple holds up to the leading x86 architectures in a variety of tasks.nimishtelang - Tuesday, November 20, 2018 - link
@Andrei Hmm, the A12 consumed 10% more Joules but had 60% more perf ... I think ARM has a ways to go here. Power usage is proportional but A12 is faster at completing the task, significantly. Normalizing for energy used as opposed to watts would be more informative, as batteries are measured in W*h = Joules not Watts.A76 is not at the level of A12 Vortex.
ZolaIII - Wednesday, November 21, 2018 - link
Actually in performance/W (in the table performance/J) it's other way around even A76 is significantly higher clocked. If you put both in the sustainable performance metrics pair of A76 will perform the same as A12 two big ones. If you put them both to sustainable optimal leakage clocks for FinFET structures (1.6~1.7GHz) A76 will end using less than a half power of an A12 big core. Apple big cores in their current state are in advantage to the ARM's A76 but not on the mobile phones allowed DTP.nimishtelang - Friday, November 23, 2018 - link
Burst performance is what matters here though: how often are you doing long sustained computation on a phone other than gaming? In that sense, 2x W for > 2x perf is a totally acceptable tradeoff to get perf/J up. I think A76 is optimizing for the wrong thing and it shows here.serendip - Tuesday, November 20, 2018 - link
A76 for midrange devices is what I'm looking forward to. A cheap 670 successor with 4x A55 and 2x A76 with a total device price of $200 sounds good. I just need something to replace my ancient SD650 which surprisingly still rocks after almost three years.ZolaIII - Wednesday, November 21, 2018 - link
The S675 is already announced. Dream on about $200 phone with it, it ain't gonna happen as their is no unhealthily competition like it whose back in the days between Xiaomi & LeEco. At 250~280$ we may see cuple models.nicolaim - Tuesday, November 20, 2018 - link
Typos:"... and at the same performance point, the Cortex A76 would use only 50% of a Cortex A75."
"... I'm expecting Qualcomm go be ..."
Lolimaster - Wednesday, November 21, 2018 - link
Doesn't seem much of a jump, next thing after A76 will be a true next gen.But hey, smartphones are already fast enough for every mobile use.
My S9 can do:
1080p60fps youtube
1080p h264/HEVC anime bd rips
As fast as it can get for browsing
Smartphones reached the point of being fast enough, sales will suffer a decline in the upcoming years unless people are stupid enough to keep being influenced by "you need a new flagship every 18 months because it's better and you need it"
iwod - Wednesday, November 21, 2018 - link
Just by looking at those scores I wish Apple would gives me a A10 / A11 Edge to Edge LCD Screen for Cheaper.halcyon - Wednesday, November 21, 2018 - link
@Andrei FrumusanuAnother excellent analysis, thank you.
Q: If you could have your own benchmark suite (and not just what is available now) , what would you test and how on current/future mobile SoCs?
Andrei Frumusanu - Wednesday, November 21, 2018 - link
I would do a PCMark on steroids.Calin - Wednesday, November 21, 2018 - link
I would have loved to see a comparison (however flawed) in performance and maybe energy efficiency against a low end laptop processor - a Celeron maybe. I loved the A12 performance comparison versus Intel cores (not full processors).I find it relevant considering that ARM is pushing for always-on, always-connected computers.
Also, how about an article about why Apple is (or isn't) working for A12-based MacBook Air and so on?
Valantar - Wednesday, November 21, 2018 - link
I read the Mate 20 Pro review, but I appreciate this article as the (very interesting!) CPU testing drowned a bit in the massive camera testing portion.I'm wondering, though, will you be testing the A12X/iPad Pro any time soon? Would be extremely interesting to see this same type of testing run there, in particular regarding power draw and efficiency.
jjj - Wednesday, November 21, 2018 - link
A bit funny as the GB issues you focus on is self-inflicted, you are the one discarding the memory and crypto score.and focusing only on Int and FP.Andrei Frumusanu - Wednesday, November 21, 2018 - link
The memory scores are broken and mean nothing much.serendip - Wednesday, November 21, 2018 - link
$200 may be a tad optimistic. $250 phones now use SD636, $300 gets you SD660 and $350+ gets you SD710. That's ridiculous product segmentation in my eyes but Qualcomm knows best... $250 for SD67x and $350+ for SD7xx makes more sense.zer0hour - Wednesday, November 21, 2018 - link
Really good article Andrei, and one that will be relevant for quite a while.Damn, I might have to buy a XS to replace my S8.
andanother - Thursday, November 22, 2018 - link
IOS apps are either written in Swift or Objective-C while the Android space is Java based.I'm wondering how much the difference in languages matters when it comes to these benchmarks. Back in the day, it mattered quite a bit which language ran on a machine as some compilers were better suited to the task than others. Unless the benchmarks were written by equally competent assembly language coders (unlikely) it's not clear where the Qualcomm and Apple chips stand .
Death666Angel - Thursday, November 22, 2018 - link
The first slide promises "laptop-class performance" for the A76. That intrigued me. All the ARM option laptops I've seen were underpowered up to now and much too costly. ARM laptops seem to compete in the high end for everything (battery life, display, build quality) but often sacrifice storage and performance. I don't think I'll ever use an ARM laptop for the next 10 years.On the other hand, this has me interested in a comparison to the Jaguar cores still used in the modern Sony and MS consoles. They might very well be less powerfull than a Cortex A76, generally speaking. What sort of upgrade path does the Switch have? Nvidia isn't doing any new SOCs, right? I doubt they'll drastically switch (hah) architectures, so something like this might do it.
Any way to make a generational article similar to this? From the first phones, first dual core phones all the way here? Probably pretty difficult with all the different OS revisions and finding working hardware.