4+1 vs 4+4

[s44]

OK, so Samsung announced the new Exynos, and to absolutely no one's surprise it's a 4 A15+4 A7 big.LITTLE setup.

The question, of course, is whether is this going to be more or less power efficient than the 4+1 all-A15 setup in Tegra4. Also, are 4 A7s more or less die space than one A15?

[TuxDave]

Originally Posted by s44

Also, are 4 A7s more or less die space than one A15?

http://forums.anandtech.com/showthread.php?t=2294334

That thread shows how insanely tiny A7 is.

[s44]

Originally Posted by TuxDave

http://forums.anandtech.com/showthread.php?t=2294334

That thread shows how insanely tiny A7 is.

Whoa.

Interestingly enough, it's just a touch more than 1/4 of an A15. So the two designs are similar on die space... wonder which will work better.

(On Android, that is. We know WinRT can't do 4+1.)

[Dari]

Your question can only be answered by what Samsung does in firmware and Android. If they're primarily gunning for marketing supremacy then the answer will obviously be no. If, however, they take the time to optimize everything, then maybe they'll be able to get a lot of efficiency out of their new setup.

[Skurge]

Whats A7 performance relative to A9?

[Saylick]

Originally Posted by Skurge

Whats A7 performance relative to A9?

Roughly 75% of A9 from what I've read on these forums.

[BenSkywalker]

..

[lopri]

Is the "+1" in Tegra 4 A15? I thought the "+1" core in Tegra 3 differ from the other 4s.

[lopri]

I don't know much about mobile chips but I read somewhere that Tegra 3 is 4 x A9 plus 1 x A_ (insert smaller number than 9, likely 7). Which does make sense if power saving is the main purpose. You could (should be able to?) simply power-gate the identical 4 cores if that's not the case. e.g., shut-down 3 cores and leave only one active when the load is light.

[stormkroe]

Originally Posted by lopri

I don't know much about mobile chips but I read somewhere that Tegra 3 is 4 x A9 plus 1 x A_ (insert smaller number than 9, likely 7). Which does make sense if power saving is the main purpose. You could (should be able to?) simply power-gate the identical 4 cores if that's not the case. e.g., shut-down 3 cores and leave only one active when the load is light.

I thought #5 was also an A9, only made on a low-leakage process that couldn't clock over 700mhz or so...

[BenSkywalker]

..

[dagamer34]

There are really too many variables here, plus the power consumption of an SoC heavily depends on how the kernel would treat using the slower or faster cores. On the Exynos 5 Octo, will the kernel choose the A7s over the A15s until CPU utilization hits 50% or 100%? Will it automatically fire up the A15s when it detects a game is being played? Will it fire up just two A15 cores at once or all 4? Like I said, too many variables.

What we actually will be able to test in the future is the implementation of the CPU governor that decides what CPUs to use and what clock speeds to run them at (which can be endlessly tweaked, especially when you think about the differences in thermal limits in a smartphone and a tablet). You will never be able to hit 8W peak power usage in a smartphone like the Exynos 5250 sometimes does in the Nexus 10.

Again, so many variables.

[grkM3]

I thought I read somewhere that the a15 needed a helper core to feed its pipeline? if Samsung did this right then when running true 4 threads the 4+4 setup should out perform the tegra 4 as it has 4 helper cores doing there thing

the a15 can be up to 24 stage pipeline depending on how you want to build it and can be as small as 16 so does anyone know how tegra a15s are built?

If Samsung went with 4 baby cores it looks like they are going full 24 deep on the pipelines and using the baby core to feed them

[djgandy]

Interesting that everyone feels the need to have another solution when using A15 cores so they can fully shut them down. Really not that good on the power usage front it seems? Beware Intel in the future maybe?

[Eug]

Originally Posted by djgandy

Beware Intel in the future maybe?

Intel-based phones are already available.

Competition is good.

[djgandy]

Originally Posted by Eug

Intel-based phones are already available.

Competition is good.

Yeah I know, I meant that Intels next gen atom may actually be a better all round processor than A15. A15 seems to be struggling on power.

[BenSkywalker]

..

[djgandy]

Originally Posted by BenSkywalker

Moderate CPU loads are the issue. Why power up a high amp core when you can get buy with a very low amp one?



Intel has other problems, they can never match the die size of the ARM cores due to their rather excessive baggage they lug around for legacy support. On desktop chips where neither die size nor power are that big of an issue, it doesn't matter. For mobile parts the 5 CPU core Tegra 4 devotes most of its die space to GPU cores and still comes in smaller then Intel's parts which have two functional CPU units and a dated and tiny GPU in comparison.

Intel's only hope is the one thing they are always better at then everyone else. Fabrication technology. Given enough of a lead in fab tech, the rest is just noise.

Legacy rubbish? If it is legacy then it has been die shrunk for at least 10 years and be a spec of dust by now. Why must you insist on spouting this absolute lie about the x86 front end? You have absolutely no empirical evidence of this. It is a complete myth.

As for size, CloverTrail comes in at about 60 mm^2 on 32nm and the GPU and CPU cores are actually a fraction of that space, around 30%. The GPU has DX 10.1 support which adds to the area of the GPU significantly.

Tegra 4 is rumored to be 80 mm^2 on 28nm (assuming the Nvidia presentation slide was not just some rendered graphic) That gives Intel a lot of space to add performance, and if they start leveraging their process advantage, which they will, even if they are slightly behind on architecture that will make up for it.

T4 doesn't support OpenGL ES 3.0 from what I remember reading a decision most likely made to save on area too, so T4 is no miracle chip in fact it could look out of date very quickly.

[sontin]

Medfield is 64mm^2, Clover Trail around 100mm^2:
http://pc.watch.impress.co.jp/video/...569/575/p3.pdf

Tegra 4 is ~80mm^2 and worlds faster than Clover Trail.

[MrX8503]

Originally Posted by djgandy

Interesting that everyone feels the need to have another solution when using A15 cores so they can fully shut them down. Really not that good on the power usage front it seems? Beware Intel in the future maybe?

A15 seems to be power hungry. Probably more suited for tablets in its current state.

Intel needs to get their parts down to 4W.

[lopri]

How about discussing the techs in subject, instead of horse bidding based on projections and conjectures..?

[djgandy]

Originally Posted by sontin

Medfield is 64mm^2, Clover Trail around 100mm^2:
http://pc.watch.impress.co.jp/video/...569/575/p3.pdf

Tegra 4 is ~80mm^2 and worlds faster than Clover Trail.

hmm you are right, I must have been reading about Clovertrail+ which drops DX10.1.

[s44]

Interesting. No matter how the CPU side turns out, it looks like Tegra4 may beat Exynos on GPU. Hard to say for sure, of course, since Exynos is a familiar SGX design while Nvidia is iterating their own stuff.

[Rick James Astley]

I would be very cautious with Nvidia's marketing. These are the same guys that mutilated a benchmark until it produced a result that favored T3 over C2D and said it would hit the market Q4 2011.

And yes, A15 is superior but to beat Apple's GPU you need be on a similar die size or on a smaller process. There is simply no way around it, without new tech, you can't be that much more efficient than Apple to pull of a similar performance from a much smaller chip.

[Bateluer]

Originally Posted by s44

Interesting. No matter how the CPU side turns out, it looks like Tegra4 may beat Exynos on GPU. Hard to say for sure, of course, since Exynos is a familiar SGX design while Nvidia is iterating their own stuff.

I thought the Exynos 4210, 4412, and 5s were using ARM designed Mali GPUs not PowerVRs?