[SeekingAlpha] Cherry Trail not ramping until March 2015

[Homeles]

14nm/10nm whatever is not going to cut it for change.

[Citation needed]

 

[krumme]

For a calculation of dies per wafer go to:
http://www.silicon-edge.co.uk/j/index.php?option=com_content&view=article&id=68

In my calculation i used 1*1mm, as thats the minimum size. lol. But its actually twice the size of an a7. So in theory you can get aprox 100.000 working a7 cores including l1 of a single 28nm wafer or 0.015 $ per core.

Thats the marginal cost (but exclude that the soc is obviously far bigger). The more fixed cost for Rockchip whatever, is the same order.

Facts and numbers. And as a7 is a very lean chip, it just shows the midrange and buttom of the market just goes for the lowest price per transistor.

As predicted by Hans de Vries years ago we therefore saw tons of quad core a7 on the market. That will not change for 2015, but they will eventually be replaced by smaller A11 on 20nm. Thats what is happening, and one of the reasons why Intel loses big time on Atom.

 

[witeken]

The economics aren't different for Intel. However, Intel has a *far* more superior CPU architecture for the low-end; Silvermont.

 

[krumme]

The economics aren't different for Intel. However, Intel has a *far* more superior CPU architecture for the low-end; Silvermont.

Yes. Silvermont is far, far superior to A7 for performance. But the problem for Intel is, they are competing for the same market. Thats why we see all the record high compensation.

Silvermont should be competing with Apple A7/A8, qq 800 series, A57/A11 and even then its a expensive solution (at least compared to the highly integrated qq solution)

One of my small kids have a Samsung 3 10.1 tablet with an 32nm Atom. Its perfectly fine for her (although the app/game loading times is for some reason abnormal high - seems like an error to me) - But it plays zoombies, youtube whatever fine.

That solution was replaced by a Samsung 4 series with a7 quad core. The a7 is probably 10% of the size of that Atom. That just shows what Intel is up against. Samsung is clearly ripping them off big time, and using them go get better deals from qq and whatever is providing them with solutions. And it show in Intel results.

Samsung clearly wants to hold that position. The same goes for Apple. They want to rip of their suppliers - like they are used to. Now for the small players like Lenovo, Asus will never be dependant on Intel on the tablet or Phone market, the same way they are on the x86 market.

 

[krumme]

I know comparing cpu perf across platforms is difficult and only a small part of the equation, but what do we know about IPC for
A8 vs. BT vs. A57 vs. beema?
for integer and fpu perf?

 

[NTMBK]

Yes. Silvermont is far, far superior to A7 for performance.

Like hell it is! A7 is miles ahead of any other mobile CPU for single threaded performance, apart from the likes of Haswell.

 

[AtenRa]

Like hell it is! A7 is miles ahead of any other mobile CPU for single threaded performance, apart from the likes of Haswell.

I believe they are talking about ARM A7 cores, not Apple A7

 

[witeken]

Like hell it is! A7 is miles ahead of any other mobile CPU for single threaded performance, apart from the likes of Haswell.

Apple A7 ain't low-end, nor publicly available.

 

[krumme]

I believe they are talking about ARM A7 cores, not Apple A7

We should all be more precise about that

What is the IPC of Apple A8 vs. BT vs. Haswell? - a rought estimate?

My take about Apple is they will go a long way to be independant. They are protecting their adaptability and their product.

What is the perspective of using Apple A8 for their notebooks? - is this realistic for A9 in a year with 14nm tech?

 

[krumme]

Apple A7 ain't low-end, nor publicly available.

Apple a7 is old tech. A8 is here on 20nm in big numbers in ip6/ip6+. Its used thats what matters.
Apple A9 on 14nm is probably here in a year - if we look at prior execution.
Amd k12 a year and a half on 14nm.

The arm ecosystem is just exploding. How many years back was it arm a11. Man its crazy. Even an ultra small, slim in order arch like arm a11 is performing faster than an arm a9 used in ip4s.

 

[Nothingness]

However, Intel has a *far* more superior CPU architecture for the low-end; Silvermont.

How things have changed, Silvermont is now a low-end CPU

Here is Geekbench of Z3745 (Silvermont) vs MT6592 (8 Cortex-A7). Silvermont is faster, but not miles ahead, except for memory where the poor A7 is completely destroyed.

 

[witeken]

Apple a7 is old tech. A8 is here on 20nm in big numbers in ip6/ip6+. Its used thats what matters.
Apple A9 on 14nm is probably here in a year - if we look at prior execution.
Amd k12 a year and a half on 14nm.

The arm ecosystem is just exploding. How many years back was it arm a11. Man its crazy. Even an ultra small, slim in order arch like arm a11 is performing faster than an arm a9 used in ip4s.

We were talking about ARM Cortex A7, not Apple.

 

[ShintaiDK]

What is the IPC of Apple A8 vs. BT vs. Haswell? - a rought estimate?

Not much different than A7 CPU wise

 

[krumme]

We were talking about ARM Cortex A7, not Apple.

I knew that all the time - and so was i , it was NTMBK who misunderstanded it. But as i wrote we should all be precise here. In the later post i meant Apple arch - as i wrote.

 

[sm625]

I would be really surprised if the 2016 macbook air is still using an intel chip.

 

[krumme]

I would be really surprised if the 2016 macbook air is still using an intel chip.

My bet is they will go for a quad core A9 on 14nm, but what about the OS is should be written for it? - shouldnt we know by now then? - or can they keep that a secret?

From a business perspective its crazy to stay on x86 when ARM side have plenty of competitors. That will secure you get the profit - and not share it with your suppliers (For what the big players go to extremes). Add Apple seems to have a fantastic solution. They just need a few cores and som freq imo.

 

[Enigmoid]

For a calculation of dies per wafer go to:
http://www.silicon-edge.co.uk/j/index.php?option=com_content&view=article&id=68

In my calculation i used 1*1mm, as thats the minimum size. lol. But its actually twice the size of an a7. So in theory you can get aprox 100.000 working a7 cores including l1 of a single 28nm wafer or 0.015 $ per core.

Thats the marginal cost (but exclude that the soc is obviously far bigger). The more fixed cost for Rockchip whatever, is the same order.

Facts and numbers. And as a7 is a very lean chip, it just shows the midrange and buttom of the market just goes for the lowest price per transistor.

As predicted by Hans de Vries years ago we therefore saw tons of quad core a7 on the market. That will not change for 2015, but they will eventually be replaced by smaller A11 on 20nm. Thats what is happening, and one of the reasons why Intel loses big time on Atom.

Uh huh, and can the A7 run without L2 cache? Or uncore? You can't make these comparisons this way.

How things have changed, Silvermont is now a low-end CPU

Here is Geekbench of Z3745 (Silvermont) vs MT6592 (8 Cortex-A7). Silvermont is faster, but not miles ahead, except for memory where the poor A7 is completely destroyed.

Picking one of the highest A7 scores vs. mid range BT.

http://browser.primatelabs.com/geekbench3/compare/960360?baseline=925192

I see silvermont with more than twice the singlethread performance and generally more MT performance.

Not much different than A7 CPU wise

Depends if you use geekbench 3 and consider it valid, remember than the test is calibrated to a SB 2.5 ghz core getting 2500 points. With 1600 points for the apple A8 @1.4 ghz geekbench IPC is higher than SB and around HW. That said that is only geekbench.

 

[NTMBK]

I believe they are talking about ARM A7 cores, not Apple A7

Argh, Apple's terrible naming strategy strikes again

 

[Nothingness]

Picking one of the highest A7 scores vs. mid range BT.

http://browser.primatelabs.com/geekbench3/compare/960360?baseline=925192

I see silvermont with more than twice the singlethread performance and generally more MT performance.

I tried to stick with Android to keep it more comparable. And HP Omni 10 isn't low-end stuff which was the point being discussed

And BTW Z3770 is not mid-range BT by far. It's top-end of BT for tablets.

 

[krumme]

Uh huh, and can the A7 run without L2 cache? Or uncore? You can't make these comparisons this way.

Its not fair - imo it is a strawman. As i explicitely wrote, its only part of the soc - where l2 and uncore is only a minor part btw (depens what we define by uncore).

And its perfectly possible to do comparisons that way because the implementations of cortex a7 in different soc is very different but the core+l1 stays the same.

I brought some technical numbers to explain why Intel is losing big time on the mobile market. Its billions. Its freaking high loss, and makes BD a financial success. The diesize of cortex a7 is - as i also explicitely wrote - just a minor explanations for that.

But i would like to hear alternative explanation for the loss that have not been written?

And more explanations that we dont see CT so fast/or in the numbers, we anticipated?

 

[krumme]

I admit to beeing risk averse regarding investments. I am a cost control freak and disprove eg. all the crazy IT investment - we want 20 times retur bs, or giving products away for free or huge discount. In my book its weak ass business, and signs something is wrong. But yeaa, i admit to be an old conservative risk averse bastard.

But that said, i also just fail to see the greater vision for BT or CT on the mobile market. What is the bigger purpose? Who benefits? And who is going to pay for it?

 

[Enigmoid]

Its not fair - imo it is a strawman. As i explicitely wrote, its only part of the soc - where l2 and uncore is only a minor part btw (depens what we define by uncore).

And its perfectly possible to do comparisons that way because the implementations of cortex a7 in different soc is very different but the core+l1 stays the same.

I brought some technical numbers to explain why Intel is losing big time on the mobile market. Its billions. Its freaking high loss, and makes BD a financial success. The diesize of cortex a7 is - as i also explicitely wrote - just a minor explanations for that.

But i would like to hear alternative explanation for the loss that have not been written?

And more explanations that we dont see CT so fast/or in the numbers, we anticipated?

Oh I understand but I diagree that this is a strawman. L2 is often just as large as the cores if not larger. A more approapriate comparison would be cores + L2. Sure different configurations of A7 may be different but they all will have likely around 256 KB - 512 KB L2 cache per core.

If you look at the dies of multicore CPUs with a large number of cores you can see that the cache and uncore take up far larger portions of the die.

For a better comparison you need to look at the soc and compare die sizes and power levels with dies of similar performance. As qualcomm said, their is a lot more in a SOC than simply the CPU core.

Look at the difference between A7 and A15

5 times bigger. Yet if the CPU was to be designed without the A15 cores it would not be nearly 5x smaller.

 

[krumme]

5 times bigger. Yet if the CPU was to be designed without the A15 cores it would not be nearly 5x smaller.

I dont disagree here. Thats why i two times wrote "the entire soc is far bigger"

And i have also wondered why the A7 is so damn popular then. But that is just the reality on the market.

 

[simboss]

5 times bigger. Yet if the CPU was to be designed without the A15 cores it would not be nearly 5x smaller.

If you were to remove the A15, you would also remove a big chunk of the interconnect, and you probably would go for lower cost and area pieces of IP around the core: the memory interfaces would be smaller, no or rather simple ISP, no or more limited modem, ...

 

[Enigmoid]

If you were to remove the A15, you would also remove a big chunk of the interconnect, and you probably would go for lower cost and area pieces of IP around the core: the memory interfaces would be smaller, no or rather simple ISP, no or more limited modem, ...

This is true but I am assuming you hold everything else constant (same GPU, modem, etc.)