Mobile chips 2017: Qualcomm, Mediatek, Exynos and More

[dark zero]

Thanks @krumme for the information and thanks to GSMArena for the benchmarks...

http://www.gsmarena.com/samsung_gal...ragon_835_benchmark_comparison-news-24606.php

Seems that the improvement on Single Core reached their limits.... time to add L3 cache on the big chips.
Also seems that Qualcomm lost their edge in everything... getting defeated in Single AND Multi Core benchs along the defeat on their GPU department is very bad news for them...

 

[stuff_me_good]

Thanks @krumme for the information and thanks to GSMArena for the benchmarks...

http://www.gsmarena.com/samsung_gal...ragon_835_benchmark_comparison-news-24606.php

Seems that the improvement on Single Core reached their limits.... time to add L3 cache on the big chips.
Also seems that Qualcomm lost their edge in everything... getting defeated in Single AND Multi Core benchs along the defeat on their GPU department is very bad news for them...

WOW... how can that apple ARM chip be almost twice as fast as the competition in single core tasks? I have not read any mobile news for couple of years, but damn. I as remember apple chip used to be significantly faster even before, but it seems like they are pulling away every new generation. How can their same architecture be so much better? Must be something else than some fancy extra cache?

 

[teejee]

WOW... how can that apple ARM chip be almost twice as fast as the competition in single core tasks? I have not read any mobile news for couple of years, but damn. I as remember apple chip used to be significantly faster even before, but it seems like they are pulling away every new generation. How can their same architecture be so much better? Must be something else than some fancy extra cache?

Well Apple have their own CPU architecure, it is not similar to any other ARM architecture.
The reason it so much faster in single thread performance is because Apple made a decision that 2 strong CPU cores are enough for smartphones. This roughly give Apple twice as much power (watt) per core. And on top of this they use much more transistors per core than the competition (increases die area/cost).

The other ARM SOCs went to four strong cores as early as they could, and had/have no power headroom for reaching same single threaded performance as Apple.

This is an excellent example of specification hysteria (more cores) that gives us worse products instead.

Sent from my LG-D855 using Tapatalk

 

[dark zero]

Well Apple have their own CPU architecure, it is not similar to any other ARM architecture.
The reason it so much faster in single thread performance is because Apple made a decision that 2 strong CPU cores are enough for smartphones. This roughly give Apple twice as much power (watt) per core. And on top of this they use much more transistors per core than the competition (increases die area/cost).

The other ARM SOCs went to four strong cores as early as they could, and had/have no power headroom for reaching same single threaded performance as Apple.

This is an excellent example of specification hysteria (more cores) that gives us worse products instead.

Sent from my LG-D855 using Tapatalk

The curious fact is that Qualcomm did it right with Kryo. However now with he Snapdragon 835 they went on he wrong path again.

 

[Lodix]

The curious fact is that Qualcomm did it right with Kryo. However now with he Snapdragon 835 they went on he wrong path again.

You insist on saying that kryo was a good architecture when it wasn't: it has lower IPC, consumes more power and is more expensive (bigger) compared to its competitors. So why you keep saying that it is good ? Even Samsung's first iteration in custom cores is better and Cortex A73 if implemented well destroys Kryo.

 

[dark zero]

You insist on saying that kryo was a good architecture when it wasn't: it has lower IPC, consumes more power and is more expensive (bigger) compared to its competitors. So why you keep saying that it is good ? Even Samsung's first iteration in custom cores is better and Cortex A73 if implemented well destroys Kryo.

I didn't said that was good. However compared to the abomination of the A57, fares better.

 

[Lodix]

I didn't said that was good. However compared to the abomination of the A57, fares better.

Because Qualcomm fucked up the Cortex A57 implementation doesn't mean it was that horrible. It has less IPC in SPECint2000 and is less efficient compared to the Cortex A57 in the Exynos 7420 while being manufactured in the 14nm LPE instead of the better 14nmLPP that uses the SD820.

 

[dark zero]

Because Qualcomm fucked up the Cortex A57 implementation doesn't mean it was that horrible. It has less IPC in SPECint2000 and is less efficient compared to the Cortex A57 in the Exynos 7420 while being manufactured in the 14nm LPE instead of the better 14nmLPP that uses the SD820.

ARM A57 was in fact the worst ARM design ever seen... Even Exynos and their 14 nm wasn't safe of the overheating of the ARM A57.

And with a current example... Look what is happening with the Nintendo Switch.

Also, why Mediatek and Huawei avoided that design?

 

[Lodix]

ARM A57 was in fact the worst ARM design ever seen... Even Exynos and their 14 nm wasn't safe of the overheating of the ARM A57.

And with a current example... Look what is happening with the Nintendo Switch.

Also, why Mediatek and Huawei avoided that design?

The Nintendo Switch is bad designed overall.

Huawei and Mediatek go for cheap designs. The Cortex A57 is big and therefore expensive. And TSMC didn't have FinFet ready in early 2015 and even if it was ready they wouldn't have used it because of the cost of leading edge process node. And they did use Cortex A57 in other products, just not on mobile SOCs. They wouldn't have used Kryo cores in 2015 either ( I think it is even bigger than A57 ).

The S6 Exynos 7420 uses 14nmLPE, the S7 SD 820 uses 14nmLPP and a heat pipe so it has a huge advantage in dissipating heat. Not a fair comparison, when 2 cores at max load in the Exynos uses 2'9W and 3'3W in the Snapdragon at relatively same frequency ~2'1GHz. So using kryo in 2015 would have been even worse. 73>A72>>M1>A57>Kryo

Read this article: http://www.anandtech.com/show/11088/hisilicon-kirin-960-performance-and-power
This is another example of a bad implementation ruining a product.

 

[dark zero]

HAHAHAH. A57 > Kryo?
Are you nuts?
A57 is by far the worst design in performance!
All the benchmarks confirms that!

BTW...

http://wccftech.com/htc-u-11-geekbench-leak/

Seems that SD 835 is a dissapoint and it shows the limits of ARM A73 (which is based on).... time for the new uARCH for ARM.

 

[gdansk]

ARM is so far behind Apple, I'm a bit concerned. A large Cortex, with at least 1.5x the performance per MHz of the A73, seems necessary. Otherwise, vendors like Samsung/Huawei will design their own cores instead to get the desired performance. ARM should care, they make most their money from processor royalties not the architectural licenses (though it is substantial as well).

 

[dark zero]

ARM is so far behind Apple, I'm a bit concerned. A large Cortex, with at least 1.5x the performance per MHz of the A73, seems necessary. Otherwise, vendors like Samsung/Huawei will design their own cores instead to get the desired performance. ARM should care, they make most their money from processor royalties not the architectural licenses (though it is substantial as well).

Actually a new performance design is mandatory for ARM. A total revamp which forces ARM to do a Conroe to reach Apple before is too late.

Consequences to being sold by Softbank... they are abandoning ARM

 

[Lodix]

HAHAHAH. A57 > Kryo?
Are you nuts?
A57 is by far the worst design in performance!
All the benchmarks confirms that!

BTW...

http://wccftech.com/htc-u-11-geekbench-leak/

Seems that SD 835 is a dissapoint and it shows the limits of ARM A73 (which is based on).... time for the new uARCH for ARM.

I bring you data, facts, arguments and your answer is "HAHAHA are you nuts?", like really? I was trying to put some light in your thread of misinformation comparing SOCs using Antutu benchmarks... But I see there is no point in having a conversation with you.

 

[dark zero]

I bring you data, facts, arguments and your answer is "HAHAHA are you nuts?", like really? I was trying to put some light in your thread of misinformation comparing SOCs using Antutu benchmarks... But I see there is no point in having a conversation with you.

And you are keeping saying that ARM stock is far better than custom ones and is not the case on ARM A57 which was a disaster. If it weren't like that Samsung won't do starting the Mongoose chip.

 

[Lodix]

And you are keeping saying that ARM stock is far better than custom ones and is not the case on ARM A57 which was a disaster. If it weren't like that Samsung won't do starting the Mongoose chip.

Marketing? Differentiation? It is not like they desing a new arch from scratch in one day it is something that take years and they started before knowing what had ARM for the same period of time. And why did Qualcomm ditch they super fancy custom cores if they are so good in favor of stock A73 ? Really I gave you the data to compare performance/power consumption/area of different architectures that corroborates what I am sayinf but you prefer to keep being foolish and believe all marketing stuff, go ahead. That's why Qualcomm renamed the Cortex A73 to Kryo 280, marketing.

And I didn't say that stock ARM cores are better than any other custom implementations. I just said that Kryo was not as good as you mean to think and that A57 was not that horrible ( for its time ). Did you even read what I wrote?

 

[dark zero]

Marketing? Differentiation? It is not like they desing a new arch from scratch in one day it is something that take years and they started before knowing what had ARM for the same period of time. And why did Qualcomm ditch they super fancy custom cores if they are so good in favor of stock A73 ? Really I gave you the data to compare performance/power consumption/area of different architectures that corroborates what I am sayinf but you prefer to keep being foolish and believe all marketing stuff, go ahead. That's why Qualcomm renamed the Cortex A73 to Kryo 280, marketing.

And I didn't say that stock ARM cores are better than any other custom implementations. I just said that Kryo was not as good as you mean to think and that A57 was not that horrible ( for its time ). Did you even read what I wrote?

I said initially that Kryo was good? Yes, but then I saw the numbers and I found that it was still decent for their time. But the ARM A57 was bad, even with Samsung 14 nm LPE. If not, why ARM A72 managed to obtain such insulting advantage compared to the predecesor and why ARM A73 didn't managed to get that advantage compared to ARM A72?

Qualcomm renamed the A73 as Kryo 280? I don't think so. They took the ARM A73 and optimized it for Microsoft Windows Emulation. That is only a based chip on ARM A73 which is different.

Similar thing with Mongoose M2. Samsung optimized it for desktop. The problem for them are the secondary cores which are the A53... Samsung really needs to create the lower power cores.

-----------------------

And now with the news... I forgot someone important... Apple... they are showing up in Geekbench this:

Source: http://www.gsmarena.com/iphone_8_allegedly_runs_geekbench_destroys_the_competition-news-24728.php

Apple literally murdered ARM and is near X86 tiers.

 

[Qwertilot]

They've got vast resources though.

 

[Nothingness]

And now with the news... I forgot someone important... Apple... they are showing up in Geekbench this:

Source: http://www.gsmarena.com/iphone_8_allegedly_runs_geekbench_destroys_the_competition-news-24728.php

Apple literally murdered ARM and is near X86 tiers.

The likelihood of this being a fake is *very* high. At least please don't pretend it's a "news".

 

[dark zero]

The likelihood of this being a fake is *very* high. At least please don't pretend it's a "news".

Considering the current Apple chip development... It might be not that fake at all.

Even more, it would be dissapointing that improvement.

 

[AMDisTheBEST]

Mediatek helio x30 is pretty mediocre compare to Qualcomm snapdragon 835. The GPU on helio is seriously out perform by even the older Adreno 530 on snapdragon 820. Its only benefit is strong multicore which doesn't usually translate into better performance. Many Chinese OEMs are axing the chips
because performance is not consistent.

Apple should just replace all the Intel chips with their own custom ARMs in their MacBooks. Their iPhone chips already catch up with the performance of their lowest MacBook air and it is less power hungry too so why not just use ARM chips?

 

[Nothingness]

Considering the current Apple chip development... It might be not that fake at all.

Even more, it would be dissapointing that improvement.

After making some computations you are indeed correct that this might be possible. My apologies

That'd represent a ~17% IPC improvement over A10. I still find the frequency increase way too high, but we'll see.

 

[Arachnotronic]

After making some computations you are indeed correct that this might be possible. My apologies

That'd represent a ~17% IPC improvement over A10. I still find the frequency increase way too high, but we'll see.

That score is about 30% better than the A10, easily within the range of typical yoy perf increases for Apple. Only "dud" in recent memory from a CPU perf improvement perspective I can think of is the A8, but that's because 20nm was poor and Apple missed their frequency targets.

 

[dark zero]

After making some computations you are indeed correct that this might be possible. My apologies

That'd represent a ~17% IPC improvement over A10. I still find the frequency increase way too high, but we'll see.

Indeed, it might look fake BTW. However it could be realistic increase from Apple. But... Seems that they are on the limits of Dual Core. Seems that is the time to use one of the two:
- Use the full Quad Core design (that means that the next iteration will be downclocked but the new clocks will be for all of them)
- Use SMT on the big cores for better efficiency

 

[Nothingness]

Seems that they are on the limits of Dual Core. Seems that is the time to use one of the two:
- Use the full Quad Core design (that means that the next iteration will be downclocked but the new clocks will be for all of them)
- Use SMT on the big cores for better efficiency

I wonder if the MT score doesn't point to either of these two improvements: it's low for 4 big cores.

 

[Andrei.]

I said initially that Kryo was good? Yes, but then I saw the numbers and I found that it was still decent for their time. But the ARM A57 was bad, even with Samsung 14 nm LPE.

A57 on 14LPE on the 7420 was both faster and more efficient than Kryo on 14LPP.

Kryo was just pure trash. It was a gigantic core which didn't perform and had to be kept down in terms of clocks to not break the power budget. You just have to look at the die shot comparison to the 835 to see the dramatic difference between Kryo and an A73, Qualcom abandoned it for good reason.

If not, why ARM A72 managed to obtain such insulting advantage compared to the predecesor

Because HiSilicon did an implementation basically drawing basically near what is possible out of the core.

and why ARM A73 didn't managed to get that advantage compared to ARM A72?

Because HiSilicon and Qualcomm both fucked up their implementations because the chips were rushed.

Qualcomm renamed the A73 as Kryo 280? I don't think so. They took the ARM A73 and optimized it for Microsoft Windows Emulation. That is only a based chip on ARM A73 which is different.

Bullshit. They certainly didn't touch anything instruction execution related because ARM clearly explained that the Built on Cortex license doesn't incorporate such changes. All what they did (if they did it) is likely change instruction window sizes. There is basically no discernible IPC difference between an A73 in the Kirin 960 and an "A73" in the SD835.

Similar thing with Mongoose M2. Samsung optimized it for desktop. The problem for them are the secondary cores which are the A53... Samsung really needs to create the lower power cores.

Again bullshit with no substance. M2 simply fixes and evolves what they did in M1, it's not "optimised for desktop".