2013 core sizes: A7-A15-Jaguar-Atom-Haswell

[Idontcare]

You were giving him the benefit of the doubt that his claim that 75% IPC only applies in theoretical situation was actually true. You weren't asking him to expand on it, you were asking him what the real IPC is. Of course maybe you were just exposing that he has no references and nothing to pull an answer from...

You presume too much regarding what I was thinking, and doing, when I drafted the post in question.

Are you familiar with the Socratic method?

The Socratic method (also known as method of elenchus, elenctic method, Socratic irony, or Socratic debate), named after the classical Greek philosopher Socrates, is a form of inquiry and debate between individuals with opposing viewpoints based on asking and answering questions to stimulate critical thinking and to illuminate ideas. It is a dialectical method, often involving an oppositional discussion in which the defense of one point of view is pitted against the defense of another; one participant may lead another to contradict himself in some way, thus strengthening the inquirer's own point.

The Socratic method is a negative method of hypothesis elimination, in that better hypotheses are found by steadily identifying and eliminating those that lead to contradictions. The Socratic method searches for general, commonly held truths that shape opinion, and scrutinizes them to determine their consistency with other beliefs. The basic form is a series of questions formulated as tests of logic and fact intended to help a person or group discover their beliefs about some topic, exploring the definitions or logoi (singular logos), seeking to characterize the general characteristics shared by various particular instances.

The extent to which this method is employed to bring out definitions implicit in the interlocutors' beliefs, or to help them further their understanding, is called the method of maieutics. Aristotle attributed to Socrates the discovery of the method of definition and induction, which he regarded as the essence of the scientific method.

 

[Haserath]

http://www.engadget.com/2012/12/11/mediatek-launches-worlds-first-quad-core-cortex-a7-soc-we-go-h/

Better numbers thanks to NTMBK.

 

[Exophase]

http://www.engadget.com/2012/12/11/mediatek-launches-worlds-first-quad-core-cortex-a7-soc-we-go-h/

Better numbers thanks to NTMBK.

Unfortunately none of those benchmarks are CPU only. Maybe if reviewers showed sub-section scores and not just aggregate ones there'd be something you could draw a real comparison with..

 

[Hans de Vries]

ok, the 75% claim of the original post is totally bogus, as it is a comparison of the theoretical capabilities, it does assume that the code does not branch, and do not wait for memory, better said , it is just not real comparison and does not represent ANY real code.

Francois Piednoel

The claim that the IPC of the A7 is ~75% of the A9 stems from Arm documentation.
http://www.arm.com/zh/products/processors/cortex-a50/cortex-a53-processor.php (click: performance tab)

Dhrystone (DMIPS/MHz)
Cortex-A5 1.6
Cortex-A7 1.9
Cortex-A9 2.5
Cortex-A53 2.3

Performance A7 A9 and A53 at 1.2 GHz:

Or were you thinking I was comparing the A7 to Atom?

 

[firewolfsm]

A7 seems to be very computationally efficient in terms of size when compared to the A9/A15. To think 128 A7 cores could fit in the same space as four Haswell cores...

 

[krumme]

Its difficult to come by Intel Atom benchmarks.
Spect Int 2000 base for A7 is 350 at 1.2Ghz (ref. Hans link), what is the similar for the Atom?
Anand have Intel early slides showing a score of 500 at 1.5Ghz (~ score of 400 at 1.2Ghz, so A7 88% of Atom score):
http://www.anandtech.com/show/3696/...00-series-the-fastest-smartphone-processor/14)

 

[alex98]

All the results of the quadcore A7 are based on the scores found on the screenshots of the Engadget's article :http://www.engadget.com/2012/12/11/mediatek-launches-worlds-first-quad-core-cortex-a7-soc-we-go-h/

Quadrant CPU
Quadcore A7@1.2Ghz : 9146
Exynos 4 Quad* @ 1Ghz : 9467
Exynos 4 Quad @ 1.2Ghz : 11395

Antutu CPU Integrer
Quadcore A7@1.2Ghz : 3429
Exynos 4 Quad @ 1Ghz : 3090
Exynos 4 Quad @ 1.2Ghz : 3711

Antutu CPU Float-Point
Quadcore A7@1.2Ghz : 2327
Exynos 4 Quad @ 1Ghz : 2199
Exynos 4 Quad @ 1.2Ghz : 2655

* Quadcore A9.

 

[Haserath]

Unfortunately none of those benchmarks are CPU only. Maybe if reviewers showed sub-section scores and not just aggregate ones there'd be something you could draw a real comparison with..

Oops, didn't really check before posting. That's not useful at all.:hmm:

 

[blackened23]

Look how little they are compared to intel . Looks like game over for intel they best pack up shop. How can they compete with that ?

Uh, the haswell performs lightyears faster than any Cortex A part?

 

[Exophase]

All the results of the quadcore A7 are based on the scores found on the screenshots of the Engadget's article :http://www.engadget.com/2012/12/11/mediatek-launches-worlds-first-quad-core-cortex-a7-soc-we-go-h/

Quadrant CPU
Quadcore A7@1.2Ghz : 9146
Exynos 4 Quad* @ 1Ghz : 9467
Exynos 4 Quad @ 1.2Ghz : 11395

Antutu CPU Integrer
Quadcore A7@1.2Ghz : 3429
Exynos 4 Quad @ 1Ghz : 3090
Exynos 4 Quad @ 1.2Ghz : 3711

Antutu CPU Float-Point
Quadcore A7@1.2Ghz : 2327
Exynos 4 Quad @ 1Ghz : 2199
Exynos 4 Quad @ 1.2Ghz : 2655

* Quadcore A9.

Thanks.. I didn't think to actually look at the pictures for more results. Where did you get the Exynos 4 numbers..? Looking at this: http://www.antutu.com/view.shtml?id=75 it looks like the big deficits are in MEM and 2D.

The results look good for Cortex-A7, better than expected, but Antutu's integer and FP tests are probably just Dhrystone and Whetstone so not much to talk about..

 

[alex98]

Thanks.. I didn't think to actually look at the pictures for more results. Where did you get the Exynos 4 numbers..?..

I used my own GS3

 

[alex98]

Looking at this: http://www.antutu.com/view.shtml?id=75 it looks like the big deficits are in MEM and 2D.
.

Antutu's trusted results are wrong most of the time.

For example, I never managed to have more than 2900 in MEM on Antutu with my GS3. Looking at my history, my average results seems to be around 2600-2700.

 

[firewolfsm]

According to

http://www.phoronix.com/scan.php?page=article&item=samsung_exynos5_dual&num=6

The 1.7 GHz A15 Dual is about 1.5-4 times slower than a Core i3-330M (2.13GHz) but 8 of them could fit in the same die size as a Haswell core. At this point, it would definitely be in Intel's favor to license ARM, and sacrifice one of their cores to include 8 A15s.

 

[Exophase]

Very good, thanks for testing then - I've always found some of the submitted scores suspicious too. AnTuTu says people are cheating the benchmarks, I wonder what that means..

Could you give the scores for the other categories on your phone? There's probably still a pretty big deficiency in the MEM score. Part of this could be the CPU's fault, but part of it is probably down to using an inferior memory controller. MediaTek is likely using one of ARM's memory controller IP blocks while I think Samsung developed their own, and their memory performance tends to be near the top for ARM SoCs (vs others like TI and Freescale who have done much worse).

 

[Exophase]

According to

http://www.phoronix.com/scan.php?page=article&item=samsung_exynos5_dual&num=6

The 1.7 GHz A15 Dual is about 1.5-4 times slower than a Core i3-330M (2.13GHz) but 8 of them could fit in the same die size as a Haswell core. At this point, it would definitely be in Intel's favor to license ARM, and sacrifice one of their cores to include 8 A15s.

The truth is, we don't even really know if the Cortex-A15 figure Hans gave is reliable. It's derived from nVidia's Tegra 4 presentations, and nVidia probably didn't use a real die shot. Hopefully the proportions are still roughly correct, but it's hard to know for sure.

Be that as it may, there's no telling what a Cortex-A15 would size in at using Intel's 32nm or even 22nm process. TSMC has very good density. I'm very curious as to what die area Cortex-A15 takes up on Samsung's 32nm and 28nm processes. We may be able to get a decent feel for the latter if a die size is reported for Exynos 5 Octa; we can estimate what the GPU blocks would need on 32nm since it's proportional to what's in A6 or A6X, but it'd naturally be smaller due to being on a smaller process. The fact that Samsung can do this SoC at all should mean that it can't be too large.

If Intel really wanted to do something asynchronous they'd put Atom cores on it (Silvermonts), even if they're larger than A15s on the same process (don't know that yet) - heterogeneous ISAs would be a nightmare to work with and would probably be poorly supported. The best you could really hope for is a device where you can run Windows and Android side by side or dual boot using the different cores. There was a Windows + Symbian device that did this with Moorestown + ARM11.

 

[alex98]

Here is my complete results but @ 1.4Ghz with stock 4.1.2

Total : 15841
RAM : 2771
CPU Integer : 4209
CPU Float-Point : 3133
2D : 1347
3D : 3631
IO : 550
SD Write : 48
SD Read : 152

 

[lopri]

I thought Exynos 5 in Nexus 10 is dual-core A15?

 

[alex98]

It's effectivly a dualcore.

 

[Exophase]

I thought Exynos 5 in Nexus 10 is dual-core A15?

Are you referring to my post?

Exynos 5 Dual, previously named Exynos 5250, is a dual core Cortex-A15. The recently announced Exynos 5 Octa has quad Cortex-A15 and quad Cortex-A7 clusters.

 

[Khato]

The truth is, we don't even really know if the Cortex-A15 figure Hans gave is reliable. It's derived from nVidia's Tegra 4 presentations, and nVidia probably didn't use a real die shot. Hopefully the proportions are still roughly correct, but it's hard to know for sure.

Be that as it may, there's no telling what a Cortex-A15 would size in at using Intel's 32nm or even 22nm process. TSMC has very good density. I'm very curious as to what die area Cortex-A15 takes up on Samsung's 32nm and 28nm processes. We may be able to get a decent feel for the latter if a die size is reported for Exynos 5 Octa

Nice guess on where we'd get a better idea of A15 die size. Granted we have to rely upon Samsung's 19mm^2 figure for the entire quad core A15 + L2 cache and their die shot being accurate, but such gives more likely numbers than the NVIDIA figures. Especially because scaling the die shot with that area figure gives the expected area of around 0.5mm^2 for the A7 cores. Going by that same pixel -> mm scaling factor it looks to be something around 3.1mm^2 for the A15 core.

 

[CTho9305]

Nice guess on where we'd get a better idea of A15 die size. Granted we have to rely upon Samsung's 19mm^2 figure for the entire quad core A15 + L2 cache and their die shot being accurate, but such gives more likely numbers than the NVIDIA figures. Especially because scaling the die shot with that area figure gives the expected area of around 0.5mm^2 for the A7 cores. Going by that same pixel -> mm scaling factor it looks to be something around 3.1mm^2 for the A15 core.

That's a much more reasonable-looking number than the one derived from the heavily-doctored nvidia "photo", given the relative architectural complexity of A15 to the other cores.

 

[IntelUser2000]

The truth is, we don't even really know if the Cortex-A15 figure Hans gave is reliable. It's derived from nVidia's Tegra 4 presentations, and nVidia probably didn't use a real die shot. Hopefully the proportions are still roughly correct, but it's hard to know for sure.

Same here.

But we know relatively accurate numbers for the Samsung. You can see it in the die pictures of Apple's Ax chips. For 45nm, the core sizes of the A9 chip is comparable to the 45nm Atom.

 

[IntelUser2000]

Its difficult to come by Intel Atom benchmarks.
Spect Int 2000 base for A7 is 350 at 1.2Ghz (ref. Hans link), what is the similar for the Atom?

650 @ 1.6GHz.

 

[IntelUser2000]

The 1.7 GHz A15 Dual is about 1.5-4 times slower than a Core i3-330M (2.13GHz) but 8 of them could fit in the same die size as a Haswell core. At this point, it would definitely be in Intel's favor to license ARM, and sacrifice one of their cores to include 8 A15s.

Not based on the recent data. Looks like even Temash has a chance to be on par with ARM on perf/watt(if you minus the fact that it lacks advanced power management of the latter).

 

[Exophase]

Not based on the recent data. Looks like even Temash has a chance to be on par with ARM on perf/watt(if you minus the fact that it lacks advanced power management of the latter).

What can you tell me about Temash's power consumption? I'm digging for data and I'm finding quad core, 1GHz + turbo 5.9W TDP but no indication of how much the CPU cores alone use, except that it'll be under 1.5W per core. Since the Cinebench test can't turbo past 1GHz while under the ~6W TDP limit, and that uses all four cores but little GPU, I'm going to guess that it's using a fairly big chunk of the power budget and at least 1W per core.

Looks like Cortex-A15 in the implementations we have any real info on is about 1.5W/core @ 1.7GHz in Exynos 5250 (Samsung 32nm) and 1.3W/core @ 1.8GHz in Exynos Octa (Samsung 28nm). That probably varies tangibly with bins. I don't think it'll need more than 1.2GHz to be competitive with 1GHz Jaguar cores, and it'll probably need substantially less than peak power to reach that.. as I've always been saying we really need actual perf/W curves to really make a valid comparison, you can't just look at peak vs peak when the peaks represent different perf levels.

Interestingly, Samsung's 28nm Cortex-A15 looks to take a similar amount of area compared to Jaguar on TSMC's 28nm (~3mm^2). The latter process is probably denser than the former, despite both having the 28nm designation..