Samsung Exynos Thread (big.LITTLE Octa-core)

[Arachnotronic]

So there was no third party measures done? No real power virus run?

Not to my knowledge, no. This would be something that I'd very much like to see from AnandTech...

 

[witeken]

Do you really have to be insulting when I ask a simple question?

Sorry, it wasn't mean as insult. I just thought that single word fitted well. And as far as I can read, it wasn't a question.

So there was no third party measures done? No real power virus run?

http://www.anandtech.com/show/7314/intel-baytrail-preview-intel-atom-z3770-tested/4

From our SoC measurements it looks like Bay Trail’s power consumption under heavy CPU load ranges from 1W - 2.5W

 

[Sweepr]

Another bit from AnandTech's article:

A15 core @ 20nm: 1.67mm²
A57 core @ 20nm: 2.05mm² (1.22x)
Quad-core A15 cluster @ 20nm: 14.50mm²
Quad-core A57 cluster @ 20nm: 15.10mm² (1.04x)

Core size increases by 22% while cluster size increases by only 4% (A57 vs A15). Chipworks estimates the size of a single Enhanced Cyclone core @ 20nm at 12.2mm², not too far from a 4-core A57 cluster. Two completelly different and interesting design approaches.

While the Anandtech writeup was great as usual, it's a little less interesting over 4 months after the phone was introduced and with the Exynos 7420 being the shiny new thing everyone is talking about.

I really really hope they can do a follow-up with the 7420 not too long after the S6 is released.

Exynos 7420 borrows a lot from Exynos 5433 (Cortex A57/A53 + Mali T760). I'm sure they could have published the test results earlier but then we wouldn't have the insightful Cortex A57/A53 and Mali T760 analysis (that's exactly what makes AnandTech different from other websites). Now that the architectural details were explained it might not take a lot of time to publish an Exynos 7420 article in the near future.

 

[coercitiv]

Also from the Bay Trail Preview:

I had Intel measure SoC power at the board level while running a single threaded Cinebench 11.5 run on the Atom Z3770 and saw a range of 800mW - 1.2W.

 

[RussianSensation]

Looking forward to the 14nm 7420. It looks beastly. Samsung S6 should also have upgraded front and back cameras and revised TouchWiz - finally!

"As for the rest of the specs, we're looking at a 5.1-inch display (Super AMOLED, most likely) with a super high resolution of 1440 x 2560 (576 ppi), 3GB of RAM, 32GB of internal storage, a 20-megapixel rear camera, and a 5-megapixel front-facing selfie snapper. In other words, most of what we thought we knew based on rumors is shaping up to be the real deal.

Lastly, and as mentioned, the Galaxy S6 Edge set a record high, 60,978 score – by far the highest we've seen. Since AnTuTu is a multi-stage test, the break-down of the score actually allows us to spot where the Exynos 7420 improved the most compared to the older Exynos 5433 – 3D objects visualization."
http://www.phonearena.com/news/The-...-breaks-all-records-and-reveals-specs_id65798

Samsung is really determined to slap Qualcomm's 810 around. If that weren't enough, 7420 is beastly in GeekBench:

"Its octa-core Exynos 7420 chipset managed to crush it with 1,492 points in the single-core, and 5,077 points in the multi-core test. To put things in perspective, this multi-core result is way higher than any other mobile device at the moment, including the iPad Air 2, the reigning champ, which scores 4,532."
http://www.phonearena.com/news/Gala...s-everything-including-the-iPad-Air-2_id65823

Glad to see AT revising testing on the biased SunSpider bench. Oddly enough, the battery scores are very favourable for iPhone 6/6+ models at AT but other sites disagree:

"How big was the difference? The Galaxy S5 had a nearly 40% longer battery life. The HTC One (M8) outlasted the iPhone by nearly 35%, even the pretty mediocre by Android standards LG G3 outscored the iPhone 6 by a healthy double-digit percent margin. The only popular phone that had worse battery life than the iPhone 6 was the Nexus 5, but that is not really a compliment as Google’s 2013 flagship is commonly known to have poor longevity."
http://www.phonearena.com/news/Livi...y-secret-life-as-a-wall-hugger-part-3_id65772

Hopefully Samsung S6 and the new 14nm Exynos chip sell better than the S5 did as we need healthy competition in the smartphone industry.

 

[Sweepr]

Curiously 32-bit Geekbench results for Exynos 7420 were uploaded:

Exynos 5433 @ 32-bit (Galaxy Note 4)
- Multi-Thread score: 4576
- Single-Thread score: 1298

Exynos 7420 @ 32-bit (Galaxy S6)
- Multi-Thread score: 5077
- Single-Thread score: 1492

Snapdragon 810 @ 64-bit (Zperia Z4)
- Multi-Thread score: 3576
- Single-Thread score: 1196

My results are a bit above average for Exynos 5433 so I'm assuming a 15-20% improvement for Exynos 7420 (sans 64-bit). That comes from higher clocks (2.1/1.5GHz vs 1.9/1.3GHz, likely higher sustained clocks too thanks to 14nm) and LPDDR4. Should be enough to beat the crowd of S810 flagships and get power consumption down at the same time.

Snapdragon S810 results still vary a lot from phone to phone, the best scores come from LG G Flex 2 (running @ 64-bit) and even then it can barely match Exynos 5433 @ 32-bit, let alone Exynos 7420.

Here's an alleged Exynos 5433 vs Exynos 7420 AnTuTu comparison:

Again slightly better performance due to higher clocks, other than Galaxy S6 looks remarkably similar in terms of specs to the Exynos variant of the Galaxy Note 4 - both sporting Cortex A57/A53 CPU cores, Mali T760 GPU, 3GB RAM (now LPDDR4), 32GB storage and a Quad HD AMOLED screen. The GPU scores suggest a straight jump to Mali T760MP8 (up from Mali T760MP6), should be very competitive with S810's Adreno 430 (up to 30% faster than Adreno 420 according to Qualcomm).

 

[RussianSensation]

It's pretty much known now that the early S6 smartphones will ship with Exynos 7420. Qualcomm even revised their revenue down $800 Mill for 2015 as they sited losing a major customer. You can just guess who that is.

Considering 810 is on 20nm and Exynos 7420 is on 14nm, I think the new Sammy chip has it in the bank in terms of perf/watt and absolute performance. With S4, the Qualcomm was the better option with Note 4, the Exynos got a slight lead but I think this time the Exynos 7420 will be the clear choice for the S6/Note 5.

 

[Arachnotronic]

I don't know if Samsung has enough 14nm capacity in place to service a launch as large as the S6 in under a month's time.

Listening to the earnings calls of several equipment suppliers suggest that the industry is still struggling with yield problems and the foundries aren't sure if they should add more 28/20 capacity or if they should add 14/16nm capacity at current yields.

 

[Andrei.]

It's about 0.93W/core at 1.8GHz/4 core, although it's only a model.

I actually measured power in this graph, so it's not merely a model as in the 20nm page: http://images.anandtech.com/doci/8718/big-cluster_575px.png

 

[Exophase]

I actually measured power in this graph, so it's not merely a model as in the 20nm page: http://images.anandtech.com/doci/8718/big-cluster_575px.png

Oh, okay. I was confused by a comment you made earlier. That is good to know.

Have I missed where it was explained what exactly is being measured? I can infer from the article that it's at a higher level than the cluster's rail, otherwise you wouldn't be talking about turning off the other cluster, or needing to do sense resistor mods to get per-rail measurements.

 

[Andrei.]

Oh, okay. I was confused by a comment you made earlier. That is good to know.

Have I missed where it was explained what exactly is being measured? I can infer from the article that it's at a higher level than the cluster's rail, otherwise you wouldn't be talking about turning off the other cluster, or needing to do sense resistor mods to get per-rail measurements.

Separate power supply replacing the battery. I know doing per-rail is the more correct way but we just don't have the resources for that.

Everything except the graph on the 20nm page is measured power, I only had to resort to that because I don't have a 5422 or 5420 device.

 

[Exophase]

Separate power supply replacing the battery. I know doing per-rail is the more correct way but we just don't have the resources for that.

Okay, so this is the entire power consumption of the phone? I mean, I'm sure you turned off everything you could (display, baseband/radios and so on), but it has to be made clear that this at the very least includes RAM, PMIC overhead, and other stuff in the SoC. RAM and PMIC overhead in particular scale with CPU load. This greatly changes the comparison with the Intel number that AFAIK was taken from the CPU rail.

 

[Andrei.]

Okay, so this is the entire power consumption of the phone? I mean, I'm sure you turned off everything you could (display, baseband/radios and so on), but it has to be made clear that this at the very least includes RAM, PMIC overhead, and other stuff in the SoC. RAM and PMIC overhead in particular scale with CPU load. This greatly changes the comparison with the Intel number that AFAIK was taken from the CPU rail.

I explained it in the article:

Power of the screen has also been subtracted via the same methodology and verified that it is accurate enough to give a pure representation of the load power of the SoC only. On both devices the GPU and display pipeline are power-gated and as such the load power should consist primarily of the regulator overheads, CPU cores, L2 caches, system memory and interconnect bus interfaces.

My measurements pretty much match what ARM did on their own via the thorough per-rail method.

 

[Sweepr]

Andrei, what direction do you think Samsung is going next (Galaxy Note 5)?
- Souped up 14nm Cortex A57 >2.1GHz + Mali T760/T860(MP10?) / PowerVR Series7
- 14nm Cortex A72 + Mali T880 before everyone else (similar to Exynos 5433 in 2014)
- Custom 14nm ARMV8-A CPU core + custom GPU (or custom CPU core + off-the-shelf Mali T760/T860 / PowerVR Series7)

 

[Andrei.]

I don't know (and wouldn't say if I did know).

I don't think we'll be seeing Series7 from Samsung. They look like they're pretty serious about their custom architecture, we'll know more at ISSCC on Feb 28.

 

[Arachnotronic]

I don't know (and wouldn't say if I did know).

I don't think we'll be seeing Series7 from Samsung. They look like they're pretty serious about their custom architecture, we'll know more at ISSCC on Feb 28.

Here's what Samsung is presenting at ISSCC...

20nm High-κ Metal-Gate Heterogeneous 64b Quad-Core 1:30 PM
CPUs and Hexa-Core GPU for High-Performance and
Energy-Efficient Mobile Application Processor

If Samsung had a 14nm apps processor in high volume production for the S6, wouldn't one reasonably expect that they would be showing it off at ISSCC?

 

[Exophase]

I explained it in the article:

Sorry, I missed that, guess I was too eager to get to the measurements and stuff. Nonetheless, that about matches what I assumed.

My measurements pretty much match what ARM did on their own via the thorough per-rail method.

Do you know what they did specifically and on what hardware?

I want to make it clear for anyone who isn't aware, power consumption measured at the input to the device does not include that lost by the convertor. It's hard to get graphs for efficiency curves for PMICs actually used in any recent mobile devices, but I'd figure 90% at best for ~3.8V -> ~1V @ 4+A.

The LPDDR3 contribution is hard to guess. Based on numbers I've seen in the past, I suspect it'll be at least 0.5W of the total (after removing the PMIC overhead), unless there is little memory activity during this test.

I can't find a reference on how the Bay Trail power numbers were obtained but given that Intel shipped out Saltwell kits with power instrumentation of the rails to reviewers I doubt they measured at the battery.

 

[Andrei.]

Here's what Samsung is presenting at ISSCC...



If Samsung had a 14nm apps processor in high volume production for the S6, wouldn't one reasonably expect that they would be showing it off at ISSCC?

They already showcased their first custom arch in 2013, ISSCC is not usually for product announcements but closed circle industry showcases. The SoC you pointed out is not the 7420.

Do you know what they did specifically and on what hardware?

Their setup is pretty industrial and thorough. Can't divulge anything more but think of what Intel did with Anand a few years back but on all power rails.

 

[Nothingness]

I explained it in the article:

My measurements pretty much match what ARM did on their own via the thorough per-rail method.

Andrei, sorry to ask again, but what kind of program did you run on the core(s) to measure power? ARM cores are typically good at clock gating unused blocks, so for instance not running any FP/SIMD code will likely result in under-estimating power consumption.

 

[Sweepr]

N910C (Exynos) is the first receiving Lollipop OTA update out of all Galaxy Note 4 versions. It's 5.0.1 and still 32-bit for now (like Andrei predicted). I still hope that later on (after the Galaxy S6 launch when the big 64-bit push starts) they will enable 64-bit.

 

[Sweepr]

Quick comparison of two Exynos 7420 Geekbench submissions (32-bit vs 64-bit):

http://browser.primatelabs.com/geekbench3/compare/1874253?baseline=1888658

Similar integer multicore scores, 12.8% better floating point multicore score @ 64-bit.

Also, new Exynos model numbers (thanks to CLARiiON @ XDA):

Galaxy Tab S 2015, Galaxy Note 5 or perhaps something else?

 

[Andrei.]

I don't think those SoC names match what's actually being released, I'm aware of at least two upcoming model that isn't any of those, unless Samsung is going full crazy on their lineup.

 

[jdubs03]

I don't think those SoC names match what's actually being released, I'm aware of at least two upcoming model that isn't any of those, unless Samsung is going full crazy on their lineup.

Anything above the 7420? Or can you not divulge?

-

Any reason why the SHA2 score/mutlicore score is lower on the right-hand side, and by such an amount? Also I would expect that the overall scores would be higher for the 64-bit submission. To see single-core score being the same is pretty strange; maybe throttling issues?

 

[Sweepr]

Now that we have Snapdragon 810 reference platform results we can make some interesting comparisons, I'll add those results to Part 2 and Part 3 of my updated Exynos 5433 performance review.

The CPU scores are not terribly impressive, I don't think Exynos 7420 will have a lot of trouble to outperform Snapdragon 810 inside an actual phone. Using GFXBench scores we can also estimate how the graphics performance of those SoCs compare:

Mali T760MP8 @ 700MHz: ~22.75 FPS

Mali T760MP8 @ 700MHz: ~49 FPS

* Assuming a 1.3x scaling for Mali T760MP8 @ 700MHz (vs Mali T760MP6 @ 700MHz). We don't know the actual clockspeeds of Exynos 7420's GPU, I'm playing safe and assuming it's the same as Exynos 5433 (700MHz) but it could be higher than 700MHz @ Samsung's new 14nm FF process.

 

[Andrei.]

Some Koreans are saying 29fps MH / 59fps Trex. Not sure how accurate that is but the same source got the GeekBench results correct before it was public too.