Apple A14 - 5 nm, 11.8 billion transistors

Page 9 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Eug

Lifer
Mar 11, 2000
23,583
996
126
If I were taking those benchmarks as legitimate, differences in the memory configuration would be my first guess. Memory can have a big impact on multicore scores and I don't think we have any information on what kinds of memory our new iDevices use.
The Macrumors article suggests multicore often runs slower when the phone is first set up as it's dedicated processes to lots of background stuff. Presumably that would be it, otherwise the A14 is slower in multicore than the A13, which seems unlikely.
FWIW, my wife's iPhone XR scores the same as the iPad Air 3. Both have A12.

However, not surprisingly, it's easier to get iPhones to throttle. One way to get throttling is to plug in the phone so it charges the battery, and making everything warm.

But yeah, if the phone was just recently set up, it could have more background processes running. That would suggest the score with the mediocre multi-core score is from a third party reviewer, and not an intentional leak by Apple.
 

eek2121

Platinum Member
Aug 2, 2005
2,904
3,906
136
My 11 Pro scores 3453 in the multicore test. Something is not right.

If I had to guess, I'd say it's because multicore boost is lower to make up for single threaded performance. Single core clocks are at 3 GHz. I doubt it's an issue with Geekbench.
 

eek2121

Platinum Member
Aug 2, 2005
2,904
3,906
136
FWIW, my wife's iPhone XR scores the same as the iPad Air 3. Both have A12.

However, not surprisingly, it's easier to get iPhones to throttle. One way to get throttling is to plug in the phone so it charges the battery, and making everything warm.

But yeah, if the phone was just recently set up, it could have more background processes running. That would suggest the score with the mediocre multi-core score is from a third party reviewer, and not an intentional leak by Apple.

The A12 has much better multicore performance than the A13.
 
Last edited:

IvanKaramazov

Member
Jun 29, 2020
56
102
66
  • Like
Reactions: Tlh97

awesomedeluxe

Member
Feb 12, 2020
69
23
41
Did a very lazy random sample of compute (ie gpu) scores for the "iPhone13" models on Geekbench vs the iPhone 11 Pro. 9173 > 7510.4; 22% performance increase. I didn't stratify the "iPhone13" entries with more or less RAM, but looking at the results I didn't see any evidence that there was a difference in performance between what are presumably the iPhone 12 and iPhone 12 Pro models.

This is substantial increase - and a great deal higher than the modest ~8% increase derived from Apple's Time Flies comparison between the A14 and A12 GPUs.

But if Apple really had a 22% increase in the performance of their GPU cores, surely they would just say so, right? I can't really wrap my head around the crappy metrics they used at both of their events. The best I can think of is:

- Geekbench's compute benchmark is more favorable to Apple than their own "real world tests," and

- The new iPhones use LPDDR5, and the increased memory bandwidth has a larger impact on GPU performance than any actual change to the core arch or frequency
 

name99

Senior member
Sep 11, 2010
404
303
136
I'm upgrading mainly because of the camera. For the SoC I'd be happy with A12... that is if it could handle the camera, but I suspect A14 has purpose built upgrades specifically for that camera.

EDIT:

Oh and I shouldn't forget: 5G and 6 GB RAM, along with 128 GB entry level.



That iPhone 12 multi-core score is terrible in comparison to the iPad Air 4 MT scores. What gives? Surely it's not running long enough to throttle.

The multi-score values are all over the place. Suggesting that they are, in a sense, not representative at all. EITHER

- Apple is deliberately stress testing various scenarios (high temperature, artificially aged batteries, ...), places where one core doesn't overload the system but 2+4 cores will AND/OR

- the parts of the system that handle scheduling are receiving final tweaks. There is obviously OS level scheduling (which has come under some redesign in Darwin over the past few years), but there is also power control that's right on the chip (Apple has boasted about this in earlier years, since the A10). It's unclear EXACTLY what this chip FW does but one can imagine various possibilities. At the most basic level it could automatically change the DVFS settings for each core depending on that core's temperature, priority, and the overall system temperature (basically like the latest version of Intel's turbo stuff).
But one can imagine more ambitious possibilities, for example tracking for each core how often the core is missing out to L3 or DRAM -- code that frequently misses can be run at a lower frequency with very little performance penalty. You can play similar games with the memory controller (dynamically reduce memory frequency if that won't much affect performance) or perhaps even dynamically modify the NoC width or frequency.

Obviously these sorts of things will have been simulated up the wazoo before fabbing the chip, but you always want to run the chip with various settings to validate that the simulations match reality.
 

Eug

Lifer
Mar 11, 2000
23,583
996
126
The multi-score values are all over the place. Suggesting that they are, in a sense, not representative at all. EITHER

- Apple is deliberately stress testing various scenarios (high temperature, artificially aged batteries, ...), places where one core doesn't overload the system but 2+4 cores will AND/OR

- the parts of the system that handle scheduling are receiving final tweaks. There is obviously OS level scheduling (which has come under some redesign in Darwin over the past few years), but there is also power control that's right on the chip (Apple has boasted about this in earlier years, since the A10). It's unclear EXACTLY what this chip FW does but one can imagine various possibilities. At the most basic level it could automatically change the DVFS settings for each core depending on that core's temperature, priority, and the overall system temperature (basically like the latest version of Intel's turbo stuff).
But one can imagine more ambitious possibilities, for example tracking for each core how often the core is missing out to L3 or DRAM -- code that frequently misses can be run at a lower frequency with very little performance penalty. You can play similar games with the memory controller (dynamically reduce memory frequency if that won't much affect performance) or perhaps even dynamically modify the NoC width or frequency.

Obviously these sorts of things will have been simulated up the wazoo before fabbing the chip, but you always want to run the chip with various settings to validate that the simulations match reality.
A12 iPad scores are pretty consistent. However, A12 iPhone scores are all over the place, even years after the A12 release. I suspect there truly is throttling in iPhones for A12 with Geekbench multi-core, and it's similar with A14.

I can consistently get A12 iPad-like scores on my wife's A12 iPhone XR, but it necessitates that I remove it from its thick case and make sure it's not charging (with a warm battery). Other people doing such benches may not do that and thus may get throttling.

Also, I think the recent avalanche of scores along with the high variation in those scores is indicative that they are NOT from Apple. They are probably from review units, but the reviews are currently embargoed. Expect the reviews to show up very soon.
 
Last edited:
  • Like
Reactions: Tlh97

IvanKaramazov

Member
Jun 29, 2020
56
102
66
This is substantial increase - and a great deal higher than the modest ~8% increase derived from Apple's Time Flies comparison between the A14 and A12 GPUs.

My guess, and it is very much just a guess, is that Apple's 30% A12 -> A14 number is based on "real-world" tests, as you suggested, that show less improvement than Geekbench. I imagine for the average customer a GPU improvement means gaming performance, rather than Metal compute benchmarks. We'll have a better idea when the A14 starts showing up in things like gfxbench Aztec Ruins, or other benchmarks more comparable to games.
 
  • Like
Reactions: awesomedeluxe

smalM

Member
Sep 9, 2019
54
54
91
iphone13, on GB4, LPDDR5 is confirmed. Moreover, it use 4*128 neon
iPhone 12 pro
There's an iPad score too, pretty much the same.

4x NEON, what did you derive that from?

All the A14 GB5 compute scores now show a range of 8815–9530 for iPhone (18 entries, mean 9226) and 12167–12577 for iPad (9 entries, mean 12415). Doesn't look like pure chance anymore.
 

awesomedeluxe

Member
Feb 12, 2020
69
23
41
My guess, and it is very much just a guess, is that Apple's 30% A12 -> A14 number is based on "real-world" tests, as you suggested, that show less improvement than Geekbench. I imagine for the average customer a GPU improvement means gaming performance, rather than Metal compute benchmarks. We'll have a better idea when the A14 starts showing up in things like gfxbench Aztec Ruins, or other benchmarks more comparable to games.
Yeah, I don't want to extrapolate too much about GPU performance on compute scores. Apple has been keen on GPGPU for awhile now and I don't doubt that they excel in that area.

There's an iPad score too, pretty much the same.

4x NEON, what did you derive that from?

All the A14 GB5 compute scores now show a range of 8815–9530 for iPhone (18 entries, mean 9226) and 12167–12577 for iPad (9 entries, mean 12415). Doesn't look like pure chance anymore.
...But I can't help but mull over this 34.5% difference a little. The memory configurations appear to be the same, right? And Apple is saying they are the same APU. So that leaves frequency as the culprit.

It's not too surprising that Apple would push GPU frequencies on the iPad Air - unlike the lightning (?) cores, I don't think their GPU clocks have ever been too aggresive, and the iPad Air should have plenty of additional thermal headroom. But it is unusual. Metal scores from the A12 in the new iPad regular are consistent with the iPhone XS; likewise with the iPad 6th gen and the iPhone 7.

I wonder if Apple pushed frequencies on the Air because they'd already tested the bionic at higher clocks in their upcoming Macs and felt comfortable with it. We could be looking at the per-core GPU performance of Apple's upcoming Macs.


EDIT: Recanting this theory because the tasks the iPad beats the iPhone on are in no way uniform. I have no idea what's going on, but maybe they aren't the same APU.
 
Last edited:

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
Outside of the people who are interested in the hardware for the sake of the hardware itself, does the average iPhone customer really care what % faster a new iPhone is over an old one? Honestly, Apple doesn't even need to give any kind of number because it's irrelevant to 99.9% of their customers, just like they don't bother to discuss how much RAM their devices contain.

Real world performance estimates are better than a benchmark for the people who do things with their phones other than run Geekbench over and over again. Of course how many people are going to push this hardware to the limits outside of a game or two? My current iPhone is over 4 years old at this point and still feels fine in day to day use. Even the battery is holding up well. The new iPhone is probably a bajillion times better, but will I ever really notice?

It does look shiny though. I'm going to sit quietly in the corner and sip my beer while I determine if I have enough tech lust to upgrade.
 
  • Like
Reactions: Tlh97

Eug

Lifer
Mar 11, 2000
23,583
996
126
Outside of the people who are interested in the hardware for the sake of the hardware itself, does the average iPhone customer really care what % faster a new iPhone is over an old one? Honestly, Apple doesn't even need to give any kind of number because it's irrelevant to 99.9% of their customers, just like they don't bother to discuss how much RAM their devices contain.

Real world performance estimates are better than a benchmark for the people who do things with their phones other than run Geekbench over and over again. Of course how many people are going to push this hardware to the limits outside of a game or two? My current iPhone is over 4 years old at this point and still feels fine in day to day use. Even the battery is holding up well. The new iPhone is probably a bajillion times better, but will I ever really notice?
I am noticing some pauses in app loading etc. with my A10 iPhone 7 Plus with 3 GB RAM, but I'm thinking it may have more to do with memory than SoC speed.

This mainly started with iOS 13, and has continued with iOS 14.

My A10X iPad Pro with 4 GB does not exhibit this behaviour, but at this point I will not buy another 4 GB device. My next iPhone will have 6 GB RAM. As you can see, I tend to keep my iPhones a very long time.

Interestingly, I haven't this behaviour much with my wife's A10 iPad 7th generation which also has 3 GB RAM, but the usage pattern is very different with different apps, and to be honest I just haven't used that iPad that much.
 
Last edited:

awesomedeluxe

Member
Feb 12, 2020
69
23
41
Also, I think the recent avalanche of scores along with the high variation in those scores is indicative that they are NOT from Apple. They are probably from review units, but the reviews are currently embargoed. Expect the reviews to show up very soon.
Prosser said the embargo will lift Tuesday for the iPhone and Wednesday for the iPad :cool:
 

name99

Senior member
Sep 11, 2010
404
303
136
The SGEMM is 90Gflops.

I'm going to call it that this implies SVE/2 on the A14.
It always seemed to only make sense to ship the Macs with SVE/2 (and establish that as an ISA baseline), but 4xNEON really seems to clinch it. If you're a designer looking at the workloads and saying "even more vector performance makes sense" why would you NOT also accommodate that via SVE/2? (Presumably as 2x256?)
 

Doug S

Platinum Member
Feb 8, 2020
2,202
3,405
136
I was looking at GB5 results and clicked the wrong thing and discovered there are two results listed for "iMacPro1,1" filed yesterday. 3.5 GHz CPU with 32 MB L3. Score is 2024 ST, 17448 MT.

According to this result was obtained running the 64 bit x86 GB5 binary - and an AMD CPU. Not sure what the heck this is...

 
  • Like
Reactions: Mopetar

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
I was looking at GB5 results and clicked the wrong thing and discovered there are two results listed for "iMacPro1,1" filed yesterday. 3.5 GHz CPU with 32 MB L3. Score is 2024 ST, 17448 MT.

According to this result was obtained running the 64 bit x86 GB5 binary - and an AMD CPU. Not sure what the heck this is...


A hackintosh most likely. It's running at 6 Ghz, so most likely using LN2 as well.
 
  • Like
Reactions: Mopetar

name99

Senior member
Sep 11, 2010
404
303
136
I was looking at GB5 results and clicked the wrong thing and discovered there are two results listed for "iMacPro1,1" filed yesterday. 3.5 GHz CPU with 32 MB L3. Score is 2024 ST, 17448 MT.

According to this result was obtained running the 64 bit x86 GB5 binary - and an AMD CPU. Not sure what the heck this is...


There are always these weird hackintosh results. Not worth worrying about...


 

kallisX

Member
Sep 29, 2016
45
39
91
having owned both ipads and iphones, no matter how fast the cpu is, apple will find a generous way to slow your equipment to molasses within 1-2years. will never buy any apple product again. updates forced on me ruined it all.