Sounds like Qualcomm is feeling the pressure of MediaTek on mid/low range. Cellphones with the 6** series should be quite interesting from a price/performance perspective.
The big question mark is the high-end. TSMC's endless delays means we're unlikely to see 16 nm FinFET this year despite all their hype. Samsung are the only ones beside Intel on 14 now, and they are trying to steal Qualcomm's lunch at the high-end, so they won't give access to fabs. GloFo could but they are behind a few quarters.
Qualcomm's Krait ARMv8 is now shrouded in mystery. Will we even see it this year? I mean Qualcomm must be hoping so, because the midrange S620 basically nullifies the S810.
The whole point of the S810 was to use ARM's stuff in order to gain time, but are we now going to see a repeat with the A72 where the high-end will be made on a lower node and that alone?
There's little competition on the high end and the 810 is not very exciting, frankly. Qualcomm might have misstepped here in favor of the new Exynos, but the high end has limited volumes as the prices are still astronomical. The volume is in the mid range. Hopefully the time for the shitty 1GB, single-core, slow-as-hell Android phones is over and even the entry level could be decent phones. If the 620 has performance comparable to the 800, that makes it a top-notch processor for a mid range phone. Maybe we can see a series of phones with a price/performance ratio comparable to the Nexus 5, which is still an excellent phone, but with an improvable battery and camera.
Samsung's "14nm" is PR bull, it's the 20nm process they have now transferred to Finfet, meaning it's a nigh equivalent process to Intel's 22nm finfet, just like TSMC's upcoming "16nm" process is.
Of course the "process size" hasn't really been a good description of actual feature size for some time, so TSMC and Samsung feel free to just call their newer processes whatever they want without any connection to reality. TSMC claims it will have a "10nm" process in 2017, but by actual silicon density it will really be equivalent to Intel's "14nm" process.
Intel's process is optimized for regular layout so they have an advantage. As foundries, TSMC and Samsung need to leave more flexibility to their customers and be less restrictive in terms of uniformity requirements and this forces them to design rules that are less dense. This has always been the case: it's only exacerbated by the smaller nodes.
In the end, it really depends what your focus is. From a pure density perspective, yes, you're completely right. From a speed vs power consumption perspective, however, it'snot so easy. Imagine what AMD's chips would consume using 16nm node instead of 28nm (~4x scaling in gate area), just to make an example. In mobile, I'd argue that power consumption (vs speed) is key, therefore Samsung's (and TSMC) offering, seems to me to be quite relevant.
Ya, posting the same sh*t over and over again is really intuitive. The same data shows Intel's 22nm is barely better than GF's 32/28nm, 32nm barely better than 45nm, 45nm worse than 65nm. So who's the phony one now?
Ya, apparently Samsung's 14nm finFET is not equivalent to Intel's 14nm process. And although Samsung's 14nm is now production ready, there are still no products on the market (it takes time to ramp up production), whereas Intel has been shipping 14nm processors for the past 6 months. Intel is still quite far ahead.
Great article! Just a minor suggestion. It would be helpful to see specs of other series as a comparison. For example, you mention that the 618/620 would be competitive with the 808, and it would be great to see their specs/features side by side. Thanks!
Really hoping the 618 and 808 get used in real phones/tablets. Will be interesting to compare the performance of those to the 620/810, and the older 800/801/805.
Will be interesting to see which layout gives "good enough" performance with the best battery life (4+4 or 2+4).
Yeah, the 618 looks really interesting, at least if it is paired with a fast GPU. Unfortunately, it seems like the 2+4 configs end up with a lower spec GPU than the 4+4 chips. IMHO, they should do a 2+4 design with an even bigger GPU. Take the silicon saved from the 2 less 'big' cores and put it into the GPU. Should be able to get an extra shader cluster or two. That, plus fast/wide memory controller, top spec ISP+video decode/encode/etc and latest gen top spec modem block would be awesome!
It's not about our needs but their needs. If 618 is too good ,the more expensive 620 sells less. The 618 could have higher clocks easily but they don't want that ... To be fair,that's how things work, noting unusual really.
Just days ago i've called people crazy for believing the A72 on 28nm rumor lol. Still we are just assuming it's on 28nm and i guess we have no idea when they started working on it, thought time to market, die size and clocks would be terrible on 28nm. During their most recent post results call they said that the new custom cores chip will be sampling in the later half of calendar 2015 so now we kinda know the entire line.
If it's A72, not A57, in the next midrange chips, makes it sound like they're planning for fireworks with their own core on the high end.
(Or the high end is mostly same deal at higher clocks and they're just holding it back because it will come out later. But this is much less fun without wild speculation.)
I honestly think they don't know yet. I think it will very much depend on yields on 14nm FinFET and whether that will be sufficient for Qualcomm's volume for the latter half of the year or not.
If it's ready, then great. Otherwise I think Qualcomm has enough space that they will make a 4+4 Cortex A72 + A53 and by that time with a maturing 20nm process they can probably ramp up the A72s to 2.3 Ghz or something and call it Snapdragon 820. If their own custom ARM v8 cores are ready, then they'll go with that of course.
Qualcomm, where the heck are your custom designed cores? You dominated over ARM's paltry designs for years, and they haven't gotten any better as the a53 and a57 benchmarks show. The market is wide open, you have, or at least had, a team dedicated to it...
The market needs someone to save it from ARM's mediocre CPU designs. And if Qualcomm doesn't step up soon then AMD and Nvidia have their own cores coming next year to do so instead.
It still blows my mind how much performance these mobile cores can pump out with only a single 32bit memory channel or dual 32bit on the high end. 14.9GB/s shared between 8 processors and a GPU. I keep wondering how much more performance you could get if you doubled the memory channels or the width of the channels( ie. 4x32bit or 2x64bit). Are these architectures so memory bandwidth efficient they don't need more or is it the huge bottleneck it appears to be? If they don't need more bandwidth could the same principles be scaled to high end GPUs and CPUs without loosing too much efficiency? I know comparing desktop processors and GPUs to mobile parts is like comparing apples to oranges but it still makes me wonder...
Now that you guys have someone who is able to selectively disable cores on a SOC, I would dearly love to see common tasks benchmarked with varying numbers of cores enabled.
It would be great to see how mature Android and common apps running under Android are when it comes to taking advantage of multiple cores.
If someone had access to a rooted Android device running a custom kernel with a bunch of CPU governors, CPU hotplugs, and I/O schedulers available, it would be really interesting to get a deep-dive on how they affect performance across various numbers of CPUs in a big.LITTLE setup. :)
Granted, it would probably take a week+ to really get the numbers correctly, but it would definitely be interesting to read.
If anandtech.com wants to ship me the hardware and software, I'd be interested in writing it up. ;)
Why so many cores? Big.Little doesn't seem to be working out to well from a power or performance point of view. I've been running Android on an i3 4020y (1.5GHz dual core, HT but no turbo) tablet lately and it beats the pants off of any ARM device I've used in the same environment. I think Apple have the right idea with regard to ARM SoCs...
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
40 Comments
Back to Article
yankeeDDL - Wednesday, February 18, 2015 - link
Sounds like Qualcomm is feeling the pressure of MediaTek on mid/low range. Cellphones with the 6** series should be quite interesting from a price/performance perspective.Mondozai - Wednesday, February 18, 2015 - link
The big question mark is the high-end. TSMC's endless delays means we're unlikely to see 16 nm FinFET this year despite all their hype. Samsung are the only ones beside Intel on 14 now, and they are trying to steal Qualcomm's lunch at the high-end, so they won't give access to fabs. GloFo could but they are behind a few quarters.Qualcomm's Krait ARMv8 is now shrouded in mystery. Will we even see it this year? I mean Qualcomm must be hoping so, because the midrange S620 basically nullifies the S810.
The whole point of the S810 was to use ARM's stuff in order to gain time, but are we now going to see a repeat with the A72 where the high-end will be made on a lower node and that alone?
yankeeDDL - Wednesday, February 18, 2015 - link
There's little competition on the high end and the 810 is not very exciting, frankly. Qualcomm might have misstepped here in favor of the new Exynos, but the high end has limited volumes as the prices are still astronomical. The volume is in the mid range. Hopefully the time for the shitty 1GB, single-core, slow-as-hell Android phones is over and even the entry level could be decent phones. If the 620 has performance comparable to the 800, that makes it a top-notch processor for a mid range phone. Maybe we can see a series of phones with a price/performance ratio comparable to the Nexus 5, which is still an excellent phone, but with an improvable battery and camera.rahul1200 - Wednesday, February 18, 2015 - link
Next announcement in line is probably Snapdragon 815 and 820 which will be the flagship chipset in entire market.Refer: http://www.stjsgadgets-portal.com/2015/02/things-y...
Frenetic Pony - Wednesday, February 18, 2015 - link
Samsung's "14nm" is PR bull, it's the 20nm process they have now transferred to Finfet, meaning it's a nigh equivalent process to Intel's 22nm finfet, just like TSMC's upcoming "16nm" process is.Of course the "process size" hasn't really been a good description of actual feature size for some time, so TSMC and Samsung feel free to just call their newer processes whatever they want without any connection to reality. TSMC claims it will have a "10nm" process in 2017, but by actual silicon density it will really be equivalent to Intel's "14nm" process.
yankeeDDL - Thursday, February 19, 2015 - link
Intel's process is optimized for regular layout so they have an advantage. As foundries, TSMC and Samsung need to leave more flexibility to their customers and be less restrictive in terms of uniformity requirements and this forces them to design rules that are less dense. This has always been the case: it's only exacerbated by the smaller nodes.In the end, it really depends what your focus is. From a pure density perspective, yes, you're completely right. From a speed vs power consumption perspective, however, it'snot so easy. Imagine what AMD's chips would consume using 16nm node instead of 28nm (~4x scaling in gate area), just to make an example.
In mobile, I'd argue that power consumption (vs speed) is key, therefore Samsung's (and TSMC) offering, seems to me to be quite relevant.
patrickjp93 - Sunday, February 22, 2015 - link
Even Samsung isn't on 14nm FF. It's only marketing. Their 14nm FF process is less dense than their 20nm planar! https://www.semiwiki.com/forum/content/3884-who-wi...levizx - Monday, February 23, 2015 - link
Ya, posting the same sh*t over and over again is really intuitive. The same data shows Intel's 22nm is barely better than GF's 32/28nm, 32nm barely better than 45nm, 45nm worse than 65nm. So who's the phony one now?dragonsqrrl - Wednesday, February 18, 2015 - link
Ya, apparently Samsung's 14nm finFET is not equivalent to Intel's 14nm process. And although Samsung's 14nm is now production ready, there are still no products on the market (it takes time to ramp up production), whereas Intel has been shipping 14nm processors for the past 6 months. Intel is still quite far ahead.dragonsqrrl - Wednesday, February 18, 2015 - link
Sorry, somehow I managed to respond to the wrong comment. Intended to respond to Mondozai.Anato - Wednesday, February 25, 2015 - link
Who gives a shit, you aren't buying process, you are buying what the chip can do and what it takes to do it.danbob999 - Wednesday, February 18, 2015 - link
too many products. As always, only 1-2 of them will see actual deployment.A5 - Wednesday, February 18, 2015 - link
Yeah, I'd be surprised to ever see the 425 or 618 in a US/EU device.Achtung_BG - Wednesday, February 18, 2015 - link
X8LTE = 9x35 produced by 20nm, and Snapdragon 425, 618 and 620 should be similar.Andrei Frumusanu - Wednesday, February 18, 2015 - link
The X7 is the 9x35, which is already on 20nm.Achtung_BG - Wednesday, February 18, 2015 - link
Four Cortex A72 it is impossible to reach 1.8GHz in smartphone 3W TDP use 28nm. FD-SOI from GloFo?saratoga4 - Wednesday, February 18, 2015 - link
I do wonder when we'll see FD-SOI SOCs in phones.psychobriggsy - Wednesday, February 18, 2015 - link
Could be Samsung/GF 14nm as well. I read rumours that Qualcomm were another defector for 14nm.xilience - Wednesday, February 18, 2015 - link
Great article! Just a minor suggestion. It would be helpful to see specs of other series as a comparison. For example, you mention that the 618/620 would be competitive with the 808, and it would be great to see their specs/features side by side. Thanks!hailey14 - Wednesday, February 18, 2015 - link
Great article. Just a minor typo:We still know know much ...
phoenix_rizzen - Wednesday, February 18, 2015 - link
Really hoping the 618 and 808 get used in real phones/tablets. Will be interesting to compare the performance of those to the 620/810, and the older 800/801/805.Will be interesting to see which layout gives "good enough" performance with the best battery life (4+4 or 2+4).
extide - Wednesday, February 18, 2015 - link
Yeah, the 618 looks really interesting, at least if it is paired with a fast GPU. Unfortunately, it seems like the 2+4 configs end up with a lower spec GPU than the 4+4 chips. IMHO, they should do a 2+4 design with an even bigger GPU. Take the silicon saved from the 2 less 'big' cores and put it into the GPU. Should be able to get an extra shader cluster or two. That, plus fast/wide memory controller, top spec ISP+video decode/encode/etc and latest gen top spec modem block would be awesome!jjj - Wednesday, February 18, 2015 - link
It's not about our needs but their needs.If 618 is too good ,the more expensive 620 sells less. The 618 could have higher clocks easily but they don't want that ...
To be fair,that's how things work, noting unusual really.
jjj - Wednesday, February 18, 2015 - link
Just days ago i've called people crazy for believing the A72 on 28nm rumor lol.Still we are just assuming it's on 28nm and i guess we have no idea when they started working on it, thought time to market, die size and clocks would be terrible on 28nm.
During their most recent post results call they said that the new custom cores chip will be sampling in the later half of calendar 2015 so now we kinda know the entire line.
twotwotwo - Wednesday, February 18, 2015 - link
If it's A72, not A57, in the next midrange chips, makes it sound like they're planning for fireworks with their own core on the high end.(Or the high end is mostly same deal at higher clocks and they're just holding it back because it will come out later. But this is much less fun without wild speculation.)
aryonoco - Wednesday, February 18, 2015 - link
I honestly think they don't know yet. I think it will very much depend on yields on 14nm FinFET and whether that will be sufficient for Qualcomm's volume for the latter half of the year or not.If it's ready, then great. Otherwise I think Qualcomm has enough space that they will make a 4+4 Cortex A72 + A53 and by that time with a maturing 20nm process they can probably ramp up the A72s to 2.3 Ghz or something and call it Snapdragon 820. If their own custom ARM v8 cores are ready, then they'll go with that of course.
rahul1200 - Wednesday, February 18, 2015 - link
Next announcement in line is probably Snapdragon 815 and 820 which will be the flagship chipset in entire market.Refer: http://www.stjsgadgets-portal.com/2015/02/things-y...
Frenetic Pony - Wednesday, February 18, 2015 - link
Qualcomm, where the heck are your custom designed cores? You dominated over ARM's paltry designs for years, and they haven't gotten any better as the a53 and a57 benchmarks show. The market is wide open, you have, or at least had, a team dedicated to it...The market needs someone to save it from ARM's mediocre CPU designs. And if Qualcomm doesn't step up soon then AMD and Nvidia have their own cores coming next year to do so instead.
Kvaern2 - Thursday, February 19, 2015 - link
They got runover by the 64 bit train when they least expected it.takeshi7 - Wednesday, February 18, 2015 - link
why does the X10 LTE modem have lower upload speed than the X8?Andrei Frumusanu - Wednesday, February 18, 2015 - link
It can't do carrier aggregation on the upload channels.Einy0 - Wednesday, February 18, 2015 - link
It still blows my mind how much performance these mobile cores can pump out with only a single 32bit memory channel or dual 32bit on the high end. 14.9GB/s shared between 8 processors and a GPU. I keep wondering how much more performance you could get if you doubled the memory channels or the width of the channels( ie. 4x32bit or 2x64bit). Are these architectures so memory bandwidth efficient they don't need more or is it the huge bottleneck it appears to be? If they don't need more bandwidth could the same principles be scaled to high end GPUs and CPUs without loosing too much efficiency? I know comparing desktop processors and GPUs to mobile parts is like comparing apples to oranges but it still makes me wonder...BillBear - Wednesday, February 18, 2015 - link
Now that you guys have someone who is able to selectively disable cores on a SOC, I would dearly love to see common tasks benchmarked with varying numbers of cores enabled.It would be great to see how mature Android and common apps running under Android are when it comes to taking advantage of multiple cores.
victorson - Thursday, February 19, 2015 - link
This.phoenix_rizzen - Thursday, February 19, 2015 - link
If someone had access to a rooted Android device running a custom kernel with a bunch of CPU governors, CPU hotplugs, and I/O schedulers available, it would be really interesting to get a deep-dive on how they affect performance across various numbers of CPUs in a big.LITTLE setup. :)Granted, it would probably take a week+ to really get the numbers correctly, but it would definitely be interesting to read.
If anandtech.com wants to ship me the hardware and software, I'd be interested in writing it up. ;)
sseemaku - Thursday, February 19, 2015 - link
I feel Qualcom can do better in naming their products. Its too confusing!phoenix_rizzen - Thursday, February 19, 2015 - link
Yeah, it would have been nice if they incremented things by 100 for the 64-bit SoCs.Snapdragon 400, 600, 800 for the 32-bit ARMv7 SoCs.
Snapdragon 500, 700, 900 for the 64-bit ARMv8 SoCs.
And then they could have switched up to the 1000s for their custom ARMv8 SoCs.
But, that would be too logical for Qualcomm. :)
Valis - Thursday, February 19, 2015 - link
When are the Snapdragon 808 based mobiles coming? I haven't seen a single one being announced, yet.Azurael - Saturday, February 21, 2015 - link
Why so many cores? Big.Little doesn't seem to be working out to well from a power or performance point of view. I've been running Android on an i3 4020y (1.5GHz dual core, HT but no turbo) tablet lately and it beats the pants off of any ARM device I've used in the same environment. I think Apple have the right idea with regard to ARM SoCs...Azurael - Saturday, February 21, 2015 - link
too well*