SemiAccurate snapped an interesting pic during CES about the upcoming LG G6 phone. The interesting part is not the choice of CPU but the reasoning behind it.
When we were going through our photos from a recent event we noticed that we captured something we probably should not have. Actually we should have grabbed it, the person who had it on their screen should not have had it in our sight. In any case we have a huge bit of exclusive news for you about what CPU, and more importantly why, the LG G6 is using the chip it is.
Yes that is a Snapdragon 821
Yup it is a Qualcomm Snapdragon 821 not the rumored snapdragon 835. Normally you would expect a flagship device to use the latest and greatest SoC out there which is the 835 right? To be pedantic but on point the 835 is announced but not actually available yet. After a bit of digging into process tech and SoC schedules, we came to the conclusion that the ‘obvious’ answer is not so obvious or feasible for that matter. If those leaping to conclusions on what should be done bothered to ask the question about what can be done, they would come to an entirely different answer.
Accompanying any device launch since the 835 was announced we’ve seen a contingent with complaints that the time machine bringing the latest and greatest chips back in time for their use isn’t working. A bit more technical knowledge will help with the time machine issue. We need to focus on what can be done in silicon now, how long it takes, and why such decisions are made. Lets take a technical look at the LG G6 and the Qualcomm 835 as an example which is actually relevant to the phone in question.
Lets start with a bit of data, announcements relevant to the Qualcomm Snapdragon 835 SoC that ‘should’ or ‘is supposed to’ be in the LG G6. The Samsung 10LPE process, that is 10nm Low Power Early, officially went into mass production mode on October 17, 2016. Exactly a month later on November 17, Qualcomm and Samsung announced that the first product on the 10nm process wasn’t an Exynos device but the Snapdragon 835.
Qualcomm went to great lengths to point this out and extol the virtues of the Samsung partnership. This wasn’t just for show however, there is a great deal of coöperation needed to make a fully functional SoC on a bleeding edge node like 10LPE. Being the first with a complex SoC on a new process is even harder to pull off coöperation or no. Keep this point in mind for later.
The start of mass production generally means wafers in and there is a ramp so don’t expect full output for several months. Also the fabs where Samsung is making the 835 don’t swap over to 10nm overnight, you start with one line and add more. It will be more than a year before any Samsung fab has a majority of 10nm output, if any ever do. By mid-2017 though, Samsung should not be production/wafer start bound on 10nm, capacity should be available for those that need it.
Then we can look at production times for wafers and devices. At the 28nm node it took about 12 weeks for a wafer to be processed, that is blanks in and fully patterned silicon out. Add a couple weeks for testing, dicing, packaging, plus likely PoP additions like memory before you have a sale-able SoC module. That brings the total to ~14 weeks from silicon wafers in to packaged SoC modules at the door of a phone OEM.
When you step down to 20/22nm you add more complexity, more metal layers, more double patterning of layers, and much more processing steps in general. These steps add time and while 14nm is just 20nm with FinFET transistors (Same BEoL with a new FEoL) for anyone but Intel, it does add yet more time. Lets call it another two weeks for the added metals and mask steps from 28nm to 14nm for an aggregate ~16 weeks from wafers in to SoCs at the door of the device OEM. That is four months.
As you might have guessed the move from 20/14nm to 10nm is another big change, not the minimal 20nm -> 14nm stutter-step. It involves much more double patterning, adds triple patterning, and likely a few more metal layers as well. All this adds up to yet more time, lets call it another two weeks bringing the total from wafers in to SoCs at customers to a grand total of 18 weeks, 16 of that is for wafer processing. This is a rough but educated estimate so lets put a +/-2 weeks on that for a range of 16-20 weeks total. From here on out we will use the 18 week number but feel free to add whatever modifiers you like.
Then there are the ramp times. Chip fabrication starts out slow and with low yields and ends up at significantly higher yields and full volumes when multiple lines are running in parallel. Intel shows it off with a bunch of non-numbered curves like this 22nm to 14nm one.
From start of production (PRQ) to mature leveling off, gains are usually about a 10% increase in yield. Ramps follow a much steeper curve but that really depends on many factors, equipment delivery and setup being a key. SemiAccurate can’t specify the ramp timings so figure in what you want for this timing. That said it is months not days or weeks.
So now we come back to the Qualcomm 835, the flagship device made on the Samsung 10LPE process in question. Production started on October 17 and devices were shown off on November 17th but those were obviously from pre-production wafers. If all goes well, Samsung should be getting production 10LPE wafers out about 16 weeks from October 17, lets call it Feb 17. After dicing, testing, packaging, DRAM/Flash additions, and the rest, we are looking at late February delivery dates for Snapdragon 835 SoCs. This is for very low volumes though, while wafer ins will have gone up significantly by this time, there is an ~4 month lag until those come out. The ones you get out at that time are the ones you put in back in October.
Back to the real world of consumer products, you can’t make a device and sell it the same day, the logistics of shipping and vendor/carrier volume requirements dictate a bit of supply build-up before a launch. The reason for this is that if you make a high demand device like the upcoming S8 or any iPhone, you will sell millions or sometimes tens of millions the first day.
Apple sells ~70M iPhones a quarter with the first week of a new device being heavily lopsided toward higher sales. You need tens of millions of units in the hands of vendors when they go on sale. This means you have to build stock up and send them out to various warehouses in advance. Most leaks and early sales info come from this process, a carrier needs to have them before they can send them to stores as well. All this adds a minimum of two weeks for the shipping and distribution plus several more for SoC/finished device stockpiling. Lets call it 4-6 weeks for the entire build to store shelf process including stockpiling.
All the leaks SemiAccurate has heard says that Samsung will not show the S8 at MWC and plan to do the reveal a few weeks later. While our sources don’t have an unveiling date, the launch is supposed to start in Korea only on April 18th. The press launches usually happen ~2 weeks before the product release to allow reviews and ‘buzz’ to filter out, so lets call it around April 1 for that show.
Now lets look again at the Snapdragon 835. Normally a premium phone SoC is revealed a few months before CES and the first products will be announced at CES or MWC. That was the case with the Snapdragon 810, their first official device was the LG G-Flex2 which came out at CES and the 820 devices debuted in Barcelona at MWC last year. This year was no exception with the 835 being shown off in the spectacular ODG R8 and R9 VR glasses at CES.
Now lets look at volume. All of the CES introductions have been for halo products, the GFlex2 and the ODG VR glasses are not high volume products. The GFlex2 was a decent phone but it was more of a headline grabbing device for the normal Gx line than a top-selling SKU. The ODG glasses, as stunning as they are, are unlikely to sell more in more than the tens of thousands of units for the first half of 2017, ignorable numbers for the cell phone world. The first and only known Snapdragon 835 phone product, the Galaxy S8, but is isn’t officially even announced yet. It is going to launch, according to SemiAccurate’s best information anyway, in Korea only. The S8 will sell in pretty big numbers but the country has a population of about 50M people so think low millions of units not tens.
Why do we bring these points up? Volume ramps. Even if the ODG are on sale now we can safely ignore their volume for the purposes of this story. Samsung usually launches their phones world-wide or at least in many high volume markets to start with, this time is only Korea, a relatively low volume market, initially. This is a big change of form for them. Why? Almost assuredly because of Snapdragon 835 supply which is entirely dependent on Samsung’s 10LPE ramp.
Lets do a bit more checking back on schedules shall we? Remember from 10LPE wafers in to SoCs at an OEMs door is an 18 week process. Add 4-6 weeks for device production, initial supply buildup, and distribution for a total of 22-24 weeks or 5.5-6 months. Since the Galaxy S8 is likely to be built-in Korea, lets take the lower range of the post-SoC production numbers, 4 weeks for a total of 5.5 months from wafer to phones on sale in Korean stores.
October 17 plus 4.5 moths for SoC production is about March 1 give or take a few days. Add 4-6 weeks to that number for production and you get the end March to mid-April or about the exact times the rumors place the launch of the S8. Almost like the production process dictates the release schedules, and in this case the production volumes are dictated by the 10LPE process ramp as well.
Normally OEMs rush a few early wafers through testing and packaging, hand deliver them to pilot production lines, and get the press samples out a few weeks quicker than production. This usually shaves two weeks off the timetables which explains why the press launches and reveals are ~2 weeks before the real volume launches. Where does that line up with our numbers? Almost perfectly if you take the 6 week time for device production, volume ramps, and distribution.
After several pages of semiconductor production timetables you have probably forgotten that this article was about the LG G6 and the SoC it uses, the Snapdragon 821 not the Snapdragon 835, and definitely not the Samsung S8. Why did we go over all that data to talk about a product that doesn’t use a 10LPE chip or the Snapdragon 835? Easy, it explains why LG chose to use the Snapdragon 821 in the G6 quite clearly.
Once again SemiAccurate’s moles are telling us that the G6 will have its official debut at MWC in late February and volume launch worldwide in March sometime. That puts the G6 more than a month ahead of the Samsung GS8’s regional launch. If you make a more fair comparison of worldwide launches, the G6 is unlikely to be Korea only at first, that gap goes up to 1.5-2 months. For a product that has about a one year lifespan, this represents a rather massive 10%+ time to market advantage for the LG G6.
Time to market isn’t everything, especially if the competition has a shinier headline spec to tempt consumers with. This puts the pressure on LG to make their G6 with a Snapdragon 835 as well. Other than launching first why didn’t they? Remember how Samsung and Qualcomm were very friendly at their joint press announcement about the Snapdragon 835 being made on Samsung’s 10LPE process?
Remember how the 810 was made at TSMC but there was no similar fanfare? The 820 and 821 are both at Samsung too but it was barely mentioned in the briefings on the chip, SemiAccurate had to directly ask where the 821 was made before Samsung was mentioned. Weeks later they are best friends. This strongly suggests that there was a deal made for PR purposes as well as supply. If Samsung was going to make the Galaxy S8 using their own Exynos processor you can be pretty sure that it, not the Snapdragon 835, would be the lead device on 10LPE.
Normally when deals like this are cut the buyer of chips, Samsung’s phone division in this case, gets preferential supply. This isn’t to say any involved party is doing something unusual, immoral, illegal, or anything else, it is just a mutually beneficial arrangement between suppliers and customer. The only odd twist is that Samsung is both the supplier and consumer as is Qualcomm, you can almost hear the banjos playing on this canoe ride. Other than that it is business as usual, remember that LG was the lead 810 customer.
So with Samsung almost guaranteed to get all the early 835 production, lets say the worldwide launch is about two weeks after Korea and that early demand is more or less caught up with four weeks after that. This means there will be enough supply of Snapdragon 835s to satisfy a second large volume supplier ~6 weeks after the Galaxy S8’s Korean launch on April 18. Math puts volume availability of the second large volume 835 device in late May sometime, lets say mid-Q2 2017. Add the above manufacturing and stockpiling weeks for the same reasons we did for Samsung’s Galaxy S8.
Does that explain why LG went with the Snapdragon 821 for the G6? If they go with the current Snapdragon device they get a 10%+ time to market advantage. Do remember that when the G6 goes on sale, the 821 is still the current Qualcomm flagship device, the 835 won’t be out in volume device until weeks later. If LG waits for the G6 they won’t simply have to wait the added ~6 weeks for 10LPE devices to be available, they will have to wait another ~6 weeks for there to be enough volume to launch an 835 device. This all assumes that LG will be the second in line for 835s, something that may or may not be the case, SemiAccurate doesn’t have any spotted moles working on this one.
That means a hypothetical G6 with an 835 has an ~20% time to market disadvantage over the Galaxy S8 with the same SoC instead of a 10% advantage. If you look at the numbers, it is closer to a 3-4 month lag if LG chooses the 835 over the 821, a deadly number when it comes to a market that exists for new, shiny things. SemiAccurate thinks LG actually chose wisely here but for more reasons than time to market.
The first is what the 835 brings to the table. It has a vastly faster modem, Cat16 GbLTE vs Cat12 at ‘only’ 600Mbps, plus lower power usage from a new process, newer cores, ARM big/little with eight cores vs four on the 821, and slightly better graphics. If you look at the current crop of high end phones, only a scarce few support the antennas for 4×4 MIMO, channel bonding, and 256-QAM modulation needed for 600Mbps transmissions.
The overwhelming majority of phones with a Cat12/600Mbps capable modem don’t have the antennas to use the spectrum. This included the Qualcomm version of the iPhone 7 which does have the modem but looks to have been hobbled on the antenna count and modulation front to keep spec parity with the Intel modem SKUs.
What this means is most current high-end phones have more radio than they physically can use. This is a shame from the perspective of someone who likes tech, but the realistic side of me says the number of carriers that support the full Cat12 600Mbps is minimal at the current time. The LG G5 which the author currently uses is both on T-Mobile and lacks the antennas for 600Mbps operation. The inner geek in me cried a bit when he found out this problem but 17.218 seconds later I was smiling again at how good the phone was. 99.99% of the population won’t have those awful 17.218 seconds, they are blissfully unaware of such things.
Short story, the lack of antennas isn’t all that important in the real world. Slightly longer short story, the lack of antennas will be a problem in a few years when Cat12 and Cat16 devices are widespread but by then this generation of phones will have been supplanted several times over. By then 5G will be the worry point and even low-end devices will do Cat12 speeds. So what does the LG G6 have for antenna diversity?
This is where things get a little hazy but SemiAccurate’s moles are saying the G6 should have the full antenna complement. In non-engineering terms the G6 should be able to do the full 600Mbps which the Snapdragon 820/821 is capable of. Even if it effectively uses the same SoC as the G5, remember the 821 is just a mid-life bump of the 820, the G6 should have a bit more performance than it’s older sibling. Some of this comes from the raw MHz bump but the real addition will be in the added LTE throughput.
There is one big caveat here though which is a big, blinking neon sign saying, “…If your carrier supports Cat12”. Most don’t but a few carriers in major markets do, T-Mobile in the US is the one making the most noise at the moment. Of course the Snapdragon 835 in the S8 will have Cat16/GbLTE so it is much faster, right? Sure if you live in Syndey and have Telstra as your carrier then data will be borderline silly fast.
From SemiAccurate’s experience Cat16 is a good thing, really good. The problem is that at the moment there is one carrier in the world offering it, or at least will be offering it really soon. By the time the S8 launches there are unlikely to be any more live deployments but a small handful of others will have launched by year’s end. In short GbLTE won’t be a performance differentiator for the overwhelming majority of consumers for at least another year because their carriers won’t have the other end lit up.
And that brings us to the last point, does theoretical performance and paper specs matter to users? You can take the view that benchmarks matter and eight cores are really necessary for some obscure use case, most sites follow this script because it makes for lots of clicks. In the real world it is really hard to show a difference for six versus eight cores, storage speed and memory channels make a much greater difference. And remember LG’s G4 and it’s 6-core SoC? Anyone think the ‘missing’ two cores were not a great trade for a better camera?
So in the end the choice LG made in using the Snapdragonn 821 over the 835 is pretty clear. The G6 with an 821 will make it to market over a quarter earlier than a hypothetical G5 with an 835. The basis for all of this is rooted in the fundamental tech of semiconductor processes, it takes time to make the damn things. That time and the conjoined volume ramp means a slow start for Samsung, backroom agreements likely ensure them preferential early supply too. LG and others will get Snapdragon 835s in 1H/2017 but by then it will almost be time for the next device, lets call it V20-next. Knowing this, a company would be foolish to wait for a chip, it makes no technical or business sense.S|A
Latest posts by Charlie Demerjian (see all)
- Where do Intel and TSMC go from here? - Aug 11, 2020
- Intel should not launch Ice Lake-SP - Aug 3, 2020
- How fast is Intel’s Ice Lake-SP CPU? - Jul 30, 2020
- What is Intel making at TSMC? - Jul 28, 2020
- Intel’s 7nm meltdown takes it’s first high level head - Jul 27, 2020