REGISTER | LOGIN
Breaking News
Blog

Apple Gives a Foundry Lesson

NO RATINGS
1 saves
View Comments: Newest First | Oldest First | Threaded View
Kevin Krewell
User Rank
Author
Re: Who's schooling whom?
Kevin Krewell   10/19/2015 2:17:02 PM
NO RATINGS
@waaromikke Actually that is exactly the point of the article. I was pointing out how and why there would be difference between two chips made by two different foundries. Lacking direct and public data from the foundries, I was reverse engineering the reasons for the differences.

While individual die will vary between wafers and even across the same wafer, test and screening can sort chips to be almost alike. But because Apple only has one spec, there's no sort and binning parts like PC processor with different speed grades. Apple has only one bin, so unless Apple wants to reject a lot of parts, it has to accept some variability. In this case, the major variable seems to be active Vcc.

IJD
User Rank
Author
Re: You could do better too
IJD   10/19/2015 11:29:12 AM
NO RATINGS
Sorry Rick, information I have is under NDA. I've seen the same info in a publically (not pubically!) available TSMC presentation (which is why I mentioned it) but I can't remember where...

http://www.tsmc.com/uploadfile/ir/quarterly/2015/16cBq/E/TSMC%201Q15%20transcript.pdf

P.8 states 16FF+ is 10% faster than 14LPP (without naming names), this means 20% lower power at the same speed (from typical cell library benchmarks).

 

rick merritt
User Rank
Author
Re: You could do better too
rick merritt   10/19/2015 11:25:43 AM
NO RATINGS
@IJD Can you give references any pubically available technical comparisons of the Samsung and TSMC 16/14nm processes?

rick merritt
User Rank
Author
Re: You could do better too
rick merritt   10/19/2015 11:25:38 AM
NO RATINGS
@IJD Can you give references any pubically available technical comparisons of the Samsung and TSMC 16/14nm processes?

dromdrom
User Rank
Author
Re: Who's schooling whom?
dromdrom   10/19/2015 7:55:28 AM
NO RATINGS
Not only the Apple web followers have little understanding. Also the EEtimes writer of this article. Two foundries from two different companies will never be able to make identical chips. Especially not if these 2 chips are made in a different process technology.

In fact even the same design made in the same fab will vary across wafers and batches.

IJD
User Rank
Author
Re: You could do better too
IJD   10/19/2015 6:20:35 AM
As well as smaller chip size Samsung are offering lower wafer pricing, so the cost difference is likely to be bigger than projected. This alone is likely to be why Apple are using Samsung as well as TSMC.

The process -- not chip -- benchmarks I have seen say that the Samsung FinFET process is ~20% higher power than TSMC for the same performance or that the TSMC process is ~10% faster for the same power -- both are due to the fact that the Samsung process needs higher voltage to get the same speed as TSMC, and most of the power when active is CV^2. Leakage is similar and therefore so is standby power, and this is the state a phone CPU spends most of its time in.


In other words, regardless of benchmarks and their inadequacy, there are very good reasons why the TSMC-equipped phones have longer battery life when the CPU is heavily loaded and taking a big chunk of the battery power -- and also why in normal operation there is no significant battery life difference. Both Geekbench and Apple are correct, but they are covering two different usage cases.

None of this should be any surprise to anyone who understands the process technology, which does seem to exclude most people -- especially Apple fans -- commenting on threads like this ;-)

realjjj
User Rank
CEO
Re: Consumer Reports surprisingly careful
realjjj   10/18/2015 6:16:23 PM
NO RATINGS
They might have better labs but the same flaws as everybody else. Just 2 chips and poor benchmarks. The article is right on that point.

They should find chips that do show a big difference in some benchmarks to better understand the issue. Properly study and measure the issue with synthetic benchmarks and then test in real world scenarios for different types of users. One page load per minute ( 11h and over 600 page loads) using just mainstream sites that are not putting much load on the SoC won't provide much helpful data. You can't observe the impact of temperature that way either. Users browse in short bursts not 1 page load per minute and 10 page loads in 1-2 minutes on a heavy site followed by 8-9 minutes of idle  would lead to very different results vs 10 loads in 10 mins on an average site. And ofc the sample size problem remains. They put minimal load on the SoC, they artificially keep the screen on more than in normal use by loading 1 page per minute so their scenario would mute any differences in SoC power consumption and is not even feasible in measuring what an average user would get since even the average user would see a bigger difference (if there is any). They don't investigate ,they don't put any effort in devising a viable methodology, they do the same thing as the others. Kinda reminds you of VW, this kind of lack of effort in testing enabled them to get away with it for so long. No idea why the world is so terrible at measuring things, from economic indicators to phone testing, it's just not good enough.

TanjB
User Rank
Author
Consumer Reports surprisingly careful
TanjB   10/18/2015 4:18:34 PM
NO RATINGS
http://www.consumerreports.org/smartphones/battery-tests-find-no-chipgate-problems-in-the-iPhone-6s

Interesting that they have a lab set up for controlled testing of cellphones.

They found no significant difference, which shows Apple have deep copy-exact capabilities despite the foundry differences.

realjjj
User Rank
CEO
You could do better too
realjjj   10/16/2015 5:58:47 PM
NO RATINGS
You assume that costs at TSMC and Samsung are equal and no reason to do so.

You state about Geekbench and Antutu "The tests did show that under heavy CPU-focused processing, the battery did drain faster". Geekbench battery life test seems to load this SoC some 25-30% and the test offers both a time and a score, the score reflects the work done while testing.  Basemark's battery test takes a similar approach providing both time and score but unlike Geekbench it does put serious load on the CPU. Geekbench had 6-8h battery life, In Basemark it's 2.5h so quite a huge difference in load. Antutu is a system benchmark, not just CPU bound and not a battery test. GFXBench is GPU centric and in its battery life test you get just under 2.5h for this device.

So you might be right about a few things but you certainly made no effort to understand the testing methodology and that's part of the game too. What's the difference on average is unclear for now and if there is a significant difference it must be mentioned that heavy users will see a greater impact so for some users the difference could be relevant.  Apple mentioned 2-3% max difference in battery life not power for the SoC so if that's for users that get 2 days battery life, the ones that get one day or less would see a much greater impact. One site even managed to find a Samsung that is better http://www.tomshardware.com/news/iphone-6s-a9-samsung-vs-tsmc,30306.html

The sad reality is that all review sites have poor to very poor methodology for phones, even the ones that try to do some tests are using questionable synthetic benchmarks so it's all rtaher pointless. To make it worse, some device makers cheat by locking clocks at max when some benchmark apps are detected. Even people that should know better make ridiculous mistakes sometimes. Recently Moor Insights had a so called study where they aim to see if 8 cores are better than 4. http://www.moorinsightsstrategy.com/research-paper-do-8-cores-really-matter-in-smartphones/  The tragedy is that they use certain Qualcomm chips that are .... thermaly chalenged and often perform better when cores are disabled. So their conclusion has minimal value and can't be generalized to other SoCs.

rick merritt
User Rank
Author
Who's schooling whom?
rick merritt   10/16/2015 4:48:21 PM
NO RATINGS
Methinks at the end of the day the Samsung vs TSMC foundry brouhaha shows how little Apple's rapid Web following understands about the nuances of modern semiconductor manufacturing.

Like Us on Facebook
EE Times on Twitter
EE Times Twitter Feed