Apple A9X the new mobile SoC king

[Rakehellion]

It's a simple test for exposing differences between silicon. It may not be what you or Apple are looking for in a test, but it is informative nonetheless.

If we get a larger sample size and the data holds up then sure there is a superior and inferior version. You may not care, but not caring doesn't make the data go away.

If you do care, then why? Because you think you're being robbed of performance or just because you're a statistician? Obviously there will be a difference of some kind because they're made by different companies.

 

[kimmel]

If you do care, then why? Because you think you're being robbed of performance or just because you're a statistician? Obviously there will be a difference of some kind because they're made by different companies.

Why do I care?

Again assuming the data is corroborated:
If you are playing a game, or as was mentioned earlier, vlc or doing anything mildly processor intensive you will most likely have less battery life simply due to the luck of the draw of your manufacturer and yet you paid the same amount of money.

Why do people run benchmarks at all? Everyone earlier in this thread was drooling over synthetic performance benchmarks and how amazing it is that Apple can get such great performance on artificial workloads. Why do people ignore data that doesn't fit their personal narrative? Good question.

No, you don't get to say, "you're running the wrong apps" or "you are holding the phone wrong". If the data is there it's there...

 

[Rakehellion]

Why do I care?

Again assuming the data is corroborated:
If you are playing a game, or as was mentioned earlier, vlc or doing anything mildly processor intensive you will most likely have less battery life simply due to the luck of the draw of your manufacturer and yet you paid the same amount of money.

Why do people run benchmarks at all? Everyone earlier in this thread was drooling over synthetic performance benchmarks and how amazing it is that Apple can get such great performance on artificial workloads. Why do people ignore data that doesn't fit their personal narrative? Good question.

No, you don't get to say, "you're running the wrong apps" or "you are holding the phone wrong". If the data is there it's there...

5% less battery life is insignificant. 70% more CPU speed is significant.
Performing calculations commonly used in programs is a useful test. Having your CPU maxed for hours on end is not a realistic test.

That's why people run benchmarks, so as not to make a mountain out of a molehill or vice versa.


But a benchmark is just a benchmark. Your Geekbench score only tells how well your hardware runs Geekbench. It's only useful in a vacuum of information.

 

[kimmel]

5% less battery life is insignificant. 70% more CPU speed is significant.
Performing calculations commonly used in programs is a useful test. Having your CPU maxed for hours on end is not a realistic test.

For the workloads I do on a phone performance has been good enough. So 70% on an artificial workload is not really significant to me. Now if you could double my battery life I would be vastly more happy. So this battery test is much more interesting to me. Maxing your cpu will help to isolate differences in said cpu apart from screen power drain. That data is still relevant and I would argue more relevant than running SHA and thinking it will make my twitter load faster.

Whatever floats your boat I guess.

 

[teejee]

5% less battery life is insignificant. 70% more CPU speed is significant.
Performing calculations commonly used in programs is a useful test. Having your CPU maxed for hours on end is not a realistic test.

That's why people run benchmarks, so as not to make a mountain out of a molehill or vice versa.


But a benchmark is just a benchmark. Your Geekbench score only tells how well your hardware runs Geekbench. It's only useful in a vacuum of information.

The duty cycle in Geekbench for Iphone6s when running the battery test is around 30% so it is no extreme load, probably similar to gaming.

 

[witeken]

Gate and channel are different things, but their length should be the same.

Thank you.

 

[stingerman]

It'll be interesting to see if Samsung's negative mobile marketing team makes fun of the Samsung A9 processor difference in the iPhone 6s...

 

[Rakehellion]

For the workloads I do on a phone performance has been good enough. So 70% on an artificial workload is not really significant to me. Now if you could double my battery life I would be vastly more happy.

Comparing TSMC to Samsung is a very different test than comparing the iPhone 6s to the iPhone 6. For identical workloads, the 6s will have better battery life than the 6 because it will race to sleep faster.

 

[Fjodor2001]

You can't have it all. Samsung 14 nm has higher transistor density than TSMC 16FF. And do we know anything about max clocks?

 

[Rakehellion]

You can't have it all. Samsung 14 nm has higher transistor density than TSMC 16FF. And do we know anything about max clocks?

They're supposedly identical in performance, but the Samsung generates more heat which gives it poorer battery life.

 

[Fjodor2001]

They're supposedly identical in performance, but the Samsung generates more heat which gives it poorer battery life.

I know. But process tech have other metrics apart from that, and often there are trade offs. E.g. you may achieve higher max clocks, but at the expense of lower transistor density, all else equal.

We already know that the Samsung A9 chip has smaller die area than the one from TSMC.

 

[Arachnotronic]

I know. But process tech have other metrics apart from that, and often there are trade offs. E.g. you may achieve higher max clocks, but at the expense of lower transistor density, all else equal.

We already know that the Samsung A9 chip has smaller die area than the one from TSMC.

The Samsung 14nm process offers a tighter gate pitch (~78nm) than the TSMC 16FF+ (~90nm) which is why the former offers slightly better density than the latter.

 

[TuxDave]

Gate and channel are different things, but their length should be the same.

I sort of disagree.

At least the way I use it, I guess there's three categories. Drawn gate length, effective gate length and effective channel length.

Drawn gate length is usually what I see during physical layout and I like to describe it as "the length at the top" since we usually look top down.

The effective gate length is usually smaller since the gates typically physically taper as you go down. Perhaps you also subtract off SD overlap while you're at it.

And then you have the all important effective channel length which is the length of the inversion channel beneath the gate. That is subject to channel length modulation and SD/gate overlap. So probably needs a * to describe the conditions when you claim this number.

 

[Thala]

The effective gate length is usually smaller since the gates typically physically taper as you go down. Perhaps you also subtract off SD overlap while you're at it.

Depends on whom you ask i guess. I read publications, which take the distance between source and drain as effective gate length L, which take the SD overlap already into account compared to the drawn gate length.
I agree, that you need to be careful with these terms. Not everyone means precisely the same when using these terms.

Regarding SD overlap, that is a parameter, which you typically try to minimize in order to reduce parasitic capacitance. At the same time SD overlap is subject to process variations.

So probably needs a * to describe the conditions when you claim this number.

Agreed.

 

[Leo9]

http://arstechnica.com/apple/2015/10/samsung-vs-tsmc-comparing-the-battery-life-of-two-apple-a9s/

 

[witeken]

So Geekbench is an exception?

 

[kpkp]

Does that look like a bug in the CPU part?

 

[Thanatosis]

I think geekbench is just not at all reflective of any kind of normal use scenario. There are very few actions that will continuously load the CPU but not the GPU and dim the screen when you normally use the phone.



I have the Samsung 6s and battery life is marginally better than my ipad air 2. If the app had told me TSMC I would've believed it, because it appears to function fine.

 

[tipoo]

http://arstechnica.com/apple/2015/10/samsung-vs-tsmc-comparing-the-battery-life-of-two-apple-a9s/

Hmm, that's interesting. It's odd that the test that touches the GPU as well as CPU has them so close, while Geekbench alone shows such a big difference. Perhaps the fabrication plant differences don't play out at the GPUs lower clock speeds?

 

[Rakehellion]

Maybe Sammy has a better GPU and TSMC has a better CPU.

 

[stingerman]

Maybe Sammy has a better GPU and TSMC has a better CPU.

They are surprisingly very similar considering two different transistor charactieristics. Pretty impressive how Apple has been able to do both implementations within such narrow metrics.

 

[asendra]

I'm no expert, but just like intel has different voltages for the same CPU to achieve X Ghz, It would seem that the Samsung CPU need also a bit higher voltage to run at full Ghz, and given that geekbench is the only test running the cpu constantly at full Ghz, the difference becomes more evident

I don't know, that's my theory. I got a TSMC anyways

 

[Hans de Vries]

The ratio TSMC:Samsung has almost completely reversed at this moment.

So after 12,000 samples it was 60:40 while today, after
150,000 samples it is moving to 42:58 ........

So much for statistics. http://demo.hiraku.tw/CPUIdentifier/

Hans

 

[carop]

So much for statistics. http://demo.hiraku.tw/CPUIdentifier/

"Lies, damned lies, and statistics".

Apparently, TSMC lost the N16 battle in 2014.

“The looming battle is really not 16nm,” said Morris Chang, chairman of TSMC, in a recent conference call. “In my mind, the 16nm battle has already been fought… My thought (is) primarily on 10nm, not so much on 16nm.”

http://semiengineering.com/whos-winning-the-finfet-foundry-race/

 

[stingerman]

"Lies, damned lies, and statistics".

Apparently, TSMC lost the N16 battle in 2014.

http://semiengineering.com/whos-winning-the-finfet-foundry-race/

That's an old article that put Global Foundries in the FinFet lead. So that tells you how accurate it was. TSMC went the more conservative route to FinFet and are on their second generation.