ARM Cortex-A17
- Thread starter DSC
- Start date
D
Deleted member 13524
Guest
Here's anandtech's scoop about it:
http://www.anandtech.com/show/7739/arm-cortex-a17
According to them, the Cortex A17 should have a similar performance/clock to A15, while being smaller and consuming less power.
http://www.anandtech.com/show/7739/arm-cortex-a17
According to them, the Cortex A17 should have a similar performance/clock to A15, while being smaller and consuming less power.
If the gains are only in the memory subsystem, AnandTech is nearly certainly wrong in assuming 4.0DMIPS/MHz given that Dhrystone fits in the L1 cache... To me this looks like a rebranded revision of the A12 (i.e. like if they decided to call the A9r4 as the A10 instead). In fact I distinctly remember ARM saying A12+A7 configurations would be possible while they now say A17 is required due to AMBA4 support, which again points to this being a rebranding of a previously planned revision of the A12.
This also makes me wonder what ARM's licensing terms are for this; surely they cannot be getting a full new license fee for their multi-license customers with such minor changes even if there is a non-negligible performance uplift?
This also makes me wonder what ARM's licensing terms are for this; surely they cannot be getting a full new license fee for their multi-license customers with such minor changes even if there is a non-negligible performance uplift?
That is one odd cpu indeed. The A12 imho was very late already as a 32bit core (after A53/A57), makes me wonder why they have to add a revised version under a new name only a half year later obsoleting the former way before it even shipped.
Also, I'm not entirely sold on the idea of big.LITTLE of that core. Aren't they a bit too close together in perf/w to warrant the additional complexity? That'll be like intel doing 2+2 atoms with 2 silvermont cores at 2 Ghz and 2 silverthorne cores at 1.5Ghz.
Oh and the obvious question is where's the 64bit version of this A12/A17 core? You'd think they'd need a A55 much more than this updated A12 core, unless A57 happens to be a lot more power efficient than A15 was. Or maybe they want to give the arch licensees a bit of time before coming out with their own competitive design...
Also, I'm not entirely sold on the idea of big.LITTLE of that core. Aren't they a bit too close together in perf/w to warrant the additional complexity? That'll be like intel doing 2+2 atoms with 2 silvermont cores at 2 Ghz and 2 silverthorne cores at 1.5Ghz
.Oh and the obvious question is where's the 64bit version of this A12/A17 core? You'd think they'd need a A55 much more than this updated A12 core, unless A57 happens to be a lot more power efficient than A15 was. Or maybe they want to give the arch licensees a bit of time before coming out with their own competitive design...
Last edited by a moderator:
Any real technical reasons why a 64bit CPU makes more sense in a smartphone than a 32bit? How much ram will even high end smartphones have in 2015?
I'd rather wait how efficiency looks like on A53 cores compared to others before I reach any conclusion. Just because QCOM was quick to dismiss the 64bit "necessity" after the A7 appeared, doesn't mean anything. They'd better chase their heels to get their 64bit custom CPU out the door for marketing purposes.
Also, I'm not entirely sold on the idea of big.LITTLE of that core. Aren't they a bit too close together in perf/w to warrant the additional complexity? That'll be like intel doing 2+2 atoms with 2 silvermont cores at 2 Ghz and 2 silverthorne cores at 1.5Ghz
If the estimates in the table in Anand's report is correct, A17 has a little over 2x the DMIPS/MHz performance over the A7 so I'm not sure if you could say they are all that close! This assumes, of course, that the A7 uses a lot less power than the A17 at similar clocks, something which ought to be the case as I understand A7 cores only take up a tiny amount of die space.
In any case, I'm guessing that a 4+4 A7/A17 chip such as the announced MT6595 wouldn't need to use the A17 cores in most general phone usage. Even if this only used say 25% less power than the A17 cores (to pluck a random figure out of the air!), it could still be well worth using the additional die space.
The MT6592 Mediatek is currently using consists of 8 A7 cores that can be clocked up to 2.0GHz. Since you can fit afaik only 4 cores into ARM MPs at a time, that one should have 2 quad A7 clusters or else what I called in the past a "little.LITTLE" scheme. The question then was also what for (outside of synthetic bullcrap); you can easily use one cluster for the simple tasks while using a relative low frequency and wake up the second cluster at relatively high frequencies for the more demanding tasks.
Now from that point and onwards the MT6595 as a successing SoC for the above 6592 makes sense for mainstream smartphones, at least if a manufacturer like Mediatek obviously doesn't want to use A15/A7 big.LITTLE configs for smartphones.
Now from that point and onwards the MT6595 as a successing SoC for the above 6592 makes sense for mainstream smartphones, at least if a manufacturer like Mediatek obviously doesn't want to use A15/A7 big.LITTLE configs for smartphones.
D
Deleted member 13524
Guest
Or they just made a chip purely for Antutu scores and marketing prowess..
Considering that 6592 devices do not burn more power than 6589 powered devices despite higher CPU frequencies, I'd say the former technical explanation makes more sense.
It does make technical sense to have two similar clusters at different frequencies but it's unfortunate that there is no shared cache hierarchy between the two clusters (only coherency). I'm not sure how much power that's going to waste compared to the benefit you're describing (probably much less but hard to say, especially when taking the inevitable scheduling/OS inefficiency into account).
BTW on A7+A12/A17, if you think the gap is too small, you can simply use more aggressive synthesis (+voltage) on the A12/A17 to increase the gap. I really don't think it's a problem, and I'll be curious to see what the A5x equivalent of A12/A17 will look like (since they've already hinted at the A12 launch that there will be one). FWIW, also quite curious to see longer-term what SMT looks like on A6x given that's when they had implied a long time ago it would arrive, and how that will work along with big.LITTLE if they stick to it.
Making sense or not, there's a couple smartphones out there already which have 3GB ram. I could definitely see someone wanting to put 4GB in there if not in 2014 definitely in 2015 which is impossible with a 32bit CPU in a useful way (unless you sacrifice some of it or you use the virtualization stuff). Besides if you believe the marketing stuff armv8 should actually increase efficiency a bit (not because it's 64bit but because the instruction set is nicer).
And, probably just as important, 32bit phones will look pretty old for marketing purposes against some cheapish 64bit ones even if the latter aren't really any good.
Chicken egg dilemma? I'd rather have Android further optimized to consume even less memory.
To avoid misunderstandings IMHO more ram and/or 64bit will become quickly a necessity (if it isn't already) for tablets, but smartphones? Yes I know 2k resolutions and beyond but it's high time for them to invest in more useful things since resolutions cannot scale forever on 4.5-6" displays.
I can't know if Anand's table is true or not; Arun doubts the A17 numbers and I see his point. But even if A17 is on say A12 level I'd rather have an A17+A7 big.LITTLE config compared to one or dual quad A53 SoC. I know the Mediatek announcement doesn't mention frequencies for the former but I heard somewhere 2.0GHz for the A17 quad and 1.7GHz for the A7 quad.
A lot depends on how soon we see chips containing A53/A57 cores.
If Mediatek (and others) can 'beat' them onto the market with A7/A17 combos by 6 to 9 months, I suppose it may well be worth their while, especially if die size is a lot smaller and therefore cheaper with relatively comparable performance.
I do wonder how the MT6595 will compare with the Snapdragon 800 series. Competitive performance or comparable performance?
One thing is for certain, mid-range phones are going to become a heck of a lot more powerful over the course of the next year in one way or another.
Edit: As for the advance to 3GB phones, it seems that the GS5 could contain this amount of RAM. That's unless this image is a good fake:
http://www.gsmarena.com/an_alleged_...galaxy_s5_box_reveals_the_specs-news-7788.php
If Mediatek (and others) can 'beat' them onto the market with A7/A17 combos by 6 to 9 months, I suppose it may well be worth their while, especially if die size is a lot smaller and therefore cheaper with relatively comparable performance.
I do wonder how the MT6595 will compare with the Snapdragon 800 series. Competitive performance or comparable performance?
One thing is for certain, mid-range phones are going to become a heck of a lot more powerful over the course of the next year in one way or another.
Edit: As for the advance to 3GB phones, it seems that the GS5 could contain this amount of RAM. That's unless this image is a good fake:
http://www.gsmarena.com/an_alleged_...galaxy_s5_box_reveals_the_specs-news-7788.php
Last edited by a moderator:
D
Deleted member 13524
Guest
One good thing about mediatek is that they're usually paired with dual-sim phones.
Yep, I've already decided I won't buy another phone without dual-SIM support. It's potentially very useful when travelling around.
Just a pity that most of the relatively cheap chinaphones with dual-SIM don't tend to support many frequency bands. It seems that the manufacturers are keen to make users pay premium prices for decent quad-band or penta-band (or hexa-band!) support.
Just a pity that most of the relatively cheap chinaphones with dual-SIM don't tend to support many frequency bands. It seems that the manufacturers are keen to make users pay premium prices for decent quad-band or penta-band (or hexa-band!) support.
I don't see A57 cores having all that easy even in high end smartphones; yes there's big.LITTLE of course but still.
No idea; I can tell you though that some stuff out there with MTK SoCs has one hell of a price/performance ratio. I paid mid last year 180 bucks for a Zopo with a MTK6589 and while I initially thought it's just going to be an expensive experiment so far I'm pleasantly surprised. If I get bored with it or want something newer at that price point I can change my smartphone every year.
Mediatek's success is mostly for the chinese white box market IMO.
Definitely.
...and what kind of CPU will it carry?
ltcommander.data
Regular
Do you not consider PAE, which current A7/A12/A15/A17 support, to be a useful way to achieve 4GB+ RAM support? Compared to a 64-bit CPU, individual processes can't use more than 4GB of RAM, but it seems unlikely there will be many smartphone apps that require more than 4GB of RAM for themselves in the next couple of years. The more common benefit of having more than 4GB of RAM is to keep more apps in memory which PAE should be able to accomplish just fine.
PAE was utterly useless for anything but some servers (and even there it was a questionable hack) on x86, I see zero reason why it would be different on arm. Especially since the 64bit cpus you need to use more memory without hacks are already available.
