Snapdragon 820 Previews

[Sweepr]

Looking like a much needed upgrade from Snapdragon 810. Let's see how it stacks up again Exynos M1 and Apple A9/A10 in actual devices next year.

Very impressive graphics performance, beats the A9 here. Now we know why Samsung chose a massive Mali-T880 MP12 for their next Exynos SoC.

To that end, then, Snapdragon 820 looks like Qualcomm has regained their orientation. Performance is improved over 810 – usually greatly so – at both the CPU and GPU level. And for what it’s worth, while we don’t have extensive temperature/clockspeed logs from the MDP/S, at no point did the device get hot to the touch or leave us with the impression that it was heavily throttling to avoid getting hot to the touch. Power consumption and especially efficiency (Performance/W) is clearly going to be important consideration on 820 after everyone’s experiences with 810, and while we’ll have to see what the retail devices are like, after what Samsung was able to do in their own transition from 20nm to 14nm FinFET, I feel it bodes well for Qualcomm as well.

www.anandtech.com/show/9837/snapdragon-820-preview

Looking at single-core performance in Geekbench confirms what we’ve already seen in our other tests: Qualcomm’s Kryo CPU core is slower than Apple’s Twister CPU, but faster than ARM’s A57. After normalizing the clock frequency, Kryo’s integer performance is 27% faster than the A57. Apple’s Twister CPU ends up being 38% faster than Kryo. Based on what we know about the architecture for both Twister and A57—and after doing some back-of-the-envelope calculations—it looks like Kryo has a single integer multiply/divide unit with a 3-cycle latency. This is very similar to the A57 and A72, which also have a single integer multiply/divide unit, but with a longer 4-cycle latency.

www.tomshardware.com/reviews/snapdragon-820-performance-preview,4389.html

Waiting for some detailed thermal throttling and extended use analysis from retail devices.

 

[Arachnotronic]

This CPU gets absolutely clobbered in SPECint by the A9, often by 2x in some subteats including the infamous 176.gcc. Qualcomm has produced a Geekbench monster but it seems to not do so well when confronted by serious CPU workloads.

 

[witeken]

Looks good. And Qualcomm won't really have any competition at the high-end for at least another year, but by then Broxton will be virutally 2 years old, so not sure who Qualcomm has to fear besides losing Samsung devices. What Intel is in the PC space, qualcomm is in the phone space.

 

[Arachnotronic]

Looks good. And Qualcomm won't really have any competition at the high-end for at least another year, but by then Broxton will be virutally 2 years old, so not sure who Qualcomm has to fear besides losing Samsung devices. What Intel is in the PC space, qualcomm is in the phone space.

I agree that Qualcomm is the top merchant SoC vendor and will probably continue to be from here on out. The only difference is that in the smartphone market there are vertically integrated vendors who are increasingly trying to roll their own.

From a total SoC perspective, Qualcomm is still ahead of its competition in Android land and will probably stay that way.

 

[Sweepr]

I think the high-clocked Cortex A72 designs by Mediatek/Huawei & others will offer some serious competition in terms of CPU performance, but I don't see them matching that graphics performance (based on the 3DMark/GFXBench scores). The biggest question is power consumption and throttling.

Looks good. And Qualcomm won't really have any competition at the high-end for at least another year, but by then Broxton will be virutally 2 years old, so not sure who Qualcomm has to fear besides losing Samsung devices. What Intel is in the PC space, qualcomm is in the phone space.

Good old Moorefield is not that far behind:

http://images.anandtech.com/graphs/graph9251/74335.png

http://images.anandtech.com/graphs/graph9837/79348.png

Let's see how Broxton stacks up next year.

 

[Nothingness]

Good old Moorefield is not that far behind:

Let's see how Broxton stacks up next year.

Did you notice there are two Kraken scores in the review? One being more than twice better at 2426. That puts Intel Atom where it should be, in the toilets :biggrin:

 

[coercitiv]

Qualcomm has produced a Geekbench monster but it seems to not do so well when confronted by serious CPU workloads.

So you would argue that Geekbench is not a serious CPU test?

 

[Arachnotronic]

So you would argue that Geekbench is not a serious CPU test?

I think GB3 has its place, but the SPECint workloads are much more complex and probably better representative of general purpose CPU capabilities than GB3 is. Note that A9 really shines in the less trivial of the GB3 subtests such as Dijkstra and Lua relative to Kryo.

Also note that A9 has probably been in mass production since probably early 2015 in support of a massive iPhone 6s/6s Plus launch.

A10 should begin ramping soon, and I expect A10 to put even more distance between itself and Kryo as a result of higher clock speeds and continued architectural improvements.

 

[Arachnotronic]

Did you notice there are two Kraken scores in the review? One being more than twice better at 2426. That puts Intel Atom where it should be, in the toilets :biggrin:

Silvermont is not in the same league as Cortex A57 class CPUs and above. An absolute disgrace on Intel's part to have let Atom lapse in competitiveness like this again.

Intel is not serious about Atom and it shows. They should license ARM CPU cores at this point.

 

[Sweepr]

Did you notice there are two Kraken scores in the review? One being more than twice better at 2426. That puts Intel Atom where it should be, in the toilets :biggrin:

Not sure about Atom, but your post certainly belongs to the toilet. Moorefield actually does very well considering it's based on a 2013 microarchitecture, it's faster than Krait (S801/S805) in most tests. It's not meant to match a top notch high end 14nm FinFET ARM design from 2016.

Based on the scores unveiled today Intel has a real chance of catching up with Qualcomm @ CPU performance next year. Whether they will deliver or not is up for debate. I'm not expecting Apple standards but just enough to keep up with the Android players and get more design wins like the ZenFone 2.

 

[Arachnotronic]

Not sure about Atom, but your comment certainly belongs to the toilet. Moorefield actually does very well considering it's based on a 2013 microarchitecture, it's faster than Krait (S801/S805) in most tests.

Based on the scores unveiled today Intel has a real chance of catching up with Qualcomm @ CPU performance next year. Sure, not up to Apple standards but good enough to keep up with the Android players.

Sweepr, that's the problem isn't it? Had Intel really grabbed the bull by the horns and delivered enhanced architectures in 2014 and 2015 following up on Silvermont I think Intel would have been in good shape.

Instead we got Airmont which was basically a trivial improvement over Silvermont and Goldmont was pushed into 2016.

At any rate with all of the high end SoC makers putting out parts with integrated modems (Samsung, Qualcomm, and HiSilicon), even if Goldmont were A72 class, Intel would be behind at the SoC level.

It's clear that Intel isn't that concerned with mobile SoCs at this point based on comments from Krzanich at a recent investor conference. Their goal is to get as much revenue as they can, but the real reason they even care about mobile at this point is so that they can penetrate the networking silicon market.

 

[Sweepr]

At any rate with all of the high end SoC makers putting out parts with integrated modems (Samsung, Qualcomm, and HiSilicon), even if Goldmont were A72 class, Intel would be behind at the SoC level.

And that's probably the biggest problem, not raw performance. But let's keep things on topic here, there's other threads to discuss Atom.

 

[Arachnotronic]

And that's probably the biggest problem, not raw performance. But let's keep things on topic here, there's other threads to discuss Atom.

Right you are

Back to Kryo!

 

[frozentundra123456]

Sweepr, that's the problem isn't it? Had Intel really grabbed the bull by the horns and delivered enhanced architectures in 2014 and 2015 following up on Silvermont I think Intel would have been in good shape.

Instead we got Airmont which was basically a trivial improvement over Silvermont and Goldmont was pushed into 2016.

At any rate with all of the high end SoC makers putting out parts with integrated modems (Samsung, Qualcomm, and HiSilicon), even if Goldmont were A72 class, Intel would be behind at the SoC level.

It's clear that Intel isn't that concerned with mobile SoCs at this point based on comments from Krzanich at a recent investor conference. Their goal is to get as much revenue as they can, but the real reason they even care about mobile at this point is so that they can penetrate the networking silicon market.

OK, will only make one comment about this, but cant resist. This reminds me of the AMD fans who say "what can you expect, Vishera is an old architecture on an inferior node." Yea but that is the point, it is the job of a company to update and bring out new, competitive products. AMD actually has a better excuse than Intel, with their very limited resources. I dont know what is intel's problem in mobile.

I thought Bay Trail was a very nice advance, but now we get 14nm delays and uninspiring at best Cherry Trail. And still delays and mediocre products integrating modems. All the while, Apple and ARM just keep bringing out better chips and more attractive devices to further cement themselves in the market.

 

[Nothingness]

Not sure about Atom, but your post certainly belongs to the toilet. Moorefield actually does very well considering it's based on a 2013 microarchitecture, it's faster than Krait (S801/S805) in most tests. It's not meant to match a top notch high end 14nm FinFET ARM design from 2016.

Agreed. But why do you want to compare Atom against 800/805 in a 820 thread? And why did you feel the need to pick the graphics showing bad results for 820 instead of the one showing it more than twice faster than Silvermont? And now tell me which of the posts deserves being trashed, yours or mine?

Based on the scores unveiled today Intel has a real chance of catching up with Qualcomm @ CPU performance next year. Whether they will deliver or not is up for debate. I'm not expecting Apple standards but just enough to keep up with the Android players and get more design wins like the ZenFone 2.

Yes, let's see if they can make their core twice faster

Back to topic, I find it odd that browser scores are twice faster with the stock browser. I wonder how their stock broser would do on other ARM devices.

 

[Sweepr]

Agreed. But why do you want to compare Atom against 800/805 in a 820 thread?

Why do you always have to make it x86 vs ARM? Might as well lock this thread right now and let you keep your boring Atom bashing monologue.

And why did you feel the need to pick the graphics showing bad results for 820 instead of the one showing it more than twice faster than Silvermont?

In case you didn't notice I also posted Octane and PCMark results (and the OP has the Kraken score you're talking about). Not that it matter because according to you javascript benchmarks are useless and WebXPRT favours Intel (still waiting for evidence on this) so it's pretty hard to come up with a comparison.

Yes, let's see if they can make their core twice faster

If they did this (2x across the board), they would end up quite a bit ahead of Kryo/A72 and too close to their own Core lineup. They just need to close the gap to secure mainstream design wins, and if they don't screw up they probably will, Kryo is looking solid but nothing spectacular.

 

[Thala]

Not that it matter because according to you javascript benchmarks are useless and WebXPRT favours Intel (still waiting for evidence on this) so it's pretty hard to come up with a comparison.

Its not only according to him, but they are actually useless. You measure the performance of the Javascript implementation more than anything else. They might have some value as long as you keep the browser version and the ISA constant and only compare within the same ISA.

Yes, let's see if they can make their core twice faster

If you think Broxton will come anywhere close to stock A72 after the Cherrytrail disaster, more power to you.
And even in the unlikely event Broxton gets close to to A72....it would be about a year late.

 

[dawheat]

I think GB3 has its place, but the SPECint workloads are much more complex and probably better representative of general purpose CPU capabilities than GB3 is. Note that A9 really shines in the less trivial of the GB3 subtests such as Dijkstra and Lua relative to Kryo.

Also note that A9 has probably been in mass production since probably early 2015 in support of a massive iPhone 6s/6s Plus launch.

A10 should begin ramping soon, and I expect A10 to put even more distance between itself and Kryo as a result of higher clock speeds and continued architectural improvements.

Eh I think that's exaggerating the timelines by quite a bit. Apple seems to have a solid 6 months on the Android flagship SOCs, but no more - A8 in the iPhone 6 was launched in September 2014, Exynos 7420 in April 2015 which more or less matched performance.

The A9 was launched in September 2015 and the S820/Exynos M1 are expected to be devices available in Q1 2016 and more or less catch up in performance. The A10 surely will jump back ahead in September 2017 as normal and so on.

 

[Sweepr]

This is going to be fun to watch:

Samsung is claiming the Exynos 8890 will provide up to 30% higher performance and 10% better power efficiency than the Exynos 7420's - although the wording is a bit vague and doesn't specify if we're talking about a pure architectural comparison or actual implementation comparison, as previous PR numbers on the Exynos 7420 also didn't quite represent the full improvements of the chipset.

Andrei previously said the Galaxy S7 will be 100% Samsung (once again).
4+4 M1+A53 cores will provide some serious MT performance (rumoured 6.908 GB3 score).

 

[Arachnotronic]

This is going to be fun to watch:



Andrei previously said the Galaxy S7 will be 100% Samsung (once again).
4+4 M1+A53 cores will provide some serious MT performance (rumoured 6.908 GB3 score).

I think he meant 100% Samsung fabbed, which SD820 is

 

[Sweepr]

I think he meant 100% Samsung fabbed, which SD820 is

Makes sense, and fits the previous rumours (both Exynos M1 and S820 being used).

 

[Andrei.]

Makes sense, and fits the previous rumours (both Exynos M1 and S820 being used).

We'll still have to see how that ends up. There's a few more weeks before things are set in stone.

As expected I'm a bit underwhelmed by Kryo, but let's see if the final devices perform the same and most importantly how power is going to be.

 

[Arachnotronic]

We'll still have to see how that ends up. There's a few more weeks before things are set in stone.

As expected I'm a bit underwhelmed by Kryo, but let's see if the final devices perform the same and most importantly how power is going to be.

Andrei, what's your take on why Kryo does so poorly in SPECint2k but does quite well in GB3?

 

[Mopetar]

From a total SoC perspective, Qualcomm is still ahead of its competition in Android land and will probably stay that way.

And that's all they really need to do. Apple isn't going to buy from Qualcomm's competitors or sell to Qualcomm's potential customers so it only matters how their chips stack up against competitors in the Android space, which at this point is basically Samsung since no one else is making high-performance SoCs for phones.

 

[Arachnotronic]

And that's all they really need to do. Apple isn't going to buy from Qualcomm's competitors or sell to Qualcomm's potential customers so it only matters how their chips stack up against competitors in the Android space, which at this point is basically Samsung since no one else is making high-performance SoCs for phones.

That's basically right, yeah.