Review Andrei's Snapdragon 888 + Exynos 2100 review

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
Surprised more people aren't talking about this one:


It's just two devices and it's early data, but I have to agree that Samsung's 5LPE isn't looking that great from any angle. It comes as no surprise that Samsung chose their own node for the Exynos 2100, but Qualcomm doing so is kind of a bad look. They've been knocked down a peg if they still can't get access to any of TSMC's 5nm nodes. That's just my first impression anyway. To me it's opportunity missed for ARM in general.
 
  • Like
Reactions: AkulaMD and Tlh97

NTMBK

Lifer
Nov 14, 2011
10,208
4,940
136
Am I the only one who finds Andrei's graphs ridiculously hard to grok? He has loads of great information, but the presentation sucks. Graph two different variables in opposing directions, for 10 different devices with hard to track colour coding.

EDIT: Specifically the graphs like this:

SPECint06.png
 

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,575
146
but Qualcomm doing so is kind of a bad look.
I mean, look at the alternatives. Apple basically own the entirely of N5 while it's still ramping up, and it's not like N7 has plenty of supply left over with all of the various players there either.

5LPE provides a significant boost to density (which QC needed thanks to the additional 5G modems etc) and should really be cheaper than N7. Even if yields are meh and the power/perf just completely sucks, it's not like QC has very many alternative options.
 
  • Like
Reactions: AkulaMD and Tlh97

KompuKare

Golden Member
Jul 28, 2009
1,012
923
136
Am I the only one who finds Andrei's graphs ridiculously hard to grok? He has loads of great information, but the presentation sucks. Graph two different variables in opposing directions, for 10 different devices with hard to track colour coding.

EDIT: Specifically the graphs like this:

SPECint06.png
Yes, finally managed to figure out how to read them but really, what is wrong with having two stacks for each chip?
Or three with a seperate perf/watt stack?

Something like this:
QOc9ar7.png


Preferably like the German sites with hover-over dynamically recalculating a % comparison.
 
Last edited:

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
I mean, look at the alternatives. Apple basically own the entirely of N5 while it's still ramping up, and it's not like N7 has plenty of supply left over with all of the various players there either.

That's why it's a bad look. Qualcomm has consistently been one of TSMC's most loyal customers. Many of Qualcomm's most-successful ARM chips have been fabbed on recent TSMC processes. Snapdragon 855 was N7, Snapdragon 865 (and 870 lol) is N7P, but Snapdragon 888 is Samsung? Qualcomm got dropped like a bad habit. I strongly suspect that Qualcomm wanted some N5 wafers and got told "no" or "you will have to wait". Even if X1 and A78 aren't major improvements over A77 (and I'm not absolutely saying that they aren't), 5LPE is really holding them back. Neither Qualcomm nor anyone else in the ARM phone SoC business can realistically compete with Apple in the ARM SoC world if Apple has the best node wrapped up all to themselves.
 
  • Like
Reactions: Tlh97

Doug S

Platinum Member
Feb 8, 2020
2,202
3,405
136
That's why it's a bad look. Qualcomm has consistently been one of TSMC's most loyal customers. Many of Qualcomm's most-successful ARM chips have been fabbed on recent TSMC processes. Snapdragon 855 was N7, Snapdragon 865 (and 870 lol) is N7P, but Snapdragon 888 is Samsung? Qualcomm got dropped like a bad habit. I strongly suspect that Qualcomm wanted some N5 wafers and got told "no" or "you will have to wait". Even if X1 and A78 aren't major improvements over A77 (and I'm not absolutely saying that they aren't), 5LPE is really holding them back. Neither Qualcomm nor anyone else in the ARM phone SoC business can realistically compete with Apple in the ARM SoC world if Apple has the best node wrapped up all to themselves.

Well what's the alternative? Should Apple give up some of its allocation (which it paid for years in advance) to Qualcomm to be nice? Even if they did, if Qualcomm didn't think they would be getting enough for their anticipated demand for SD888 it wouldn't be any help to them. The real fix is for Samsung to catch up to TSMC - maybe if Qualcomm paid them in advance like Apple to help them build and outfit their fabs and develop their new nodes they'd be in a better competitive position?

This is perhaps an outgrowth of the limited supply of EUV machines and the fact that most layers are being made with EUV for both TSMC's and Samsung's 5nm. ASML can only produce them so fast, and the problem doesn't look likely to go away since smaller processes will need improved EUV scanners with a high numerical aperture. With so few fabs on the leading edge where EUV lives, there doesn't seem to be room for a second tool vendor to compete with ASML, so this may be the new normal.
 

blckgrffn

Diamond Member
May 1, 2003
9,110
3,029
136
www.teamjuchems.com
Well what's the alternative? Should Apple give up some of its allocation (which it paid for years in advance) to Qualcomm to be nice? Even if they did, if Qualcomm didn't think they would be getting enough for their anticipated demand for SD888 it wouldn't be any help to them. The real fix is for Samsung to catch up to TSMC - maybe if Qualcomm paid them in advance like Apple to help them build and outfit their fabs and develop their new nodes they'd be in a better competitive position?

This is perhaps an outgrowth of the limited supply of EUV machines and the fact that most layers are being made with EUV for both TSMC's and Samsung's 5nm. ASML can only produce them so fast, and the problem doesn't look likely to go away since smaller processes will need improved EUV scanners with a high numerical aperture. With so few fabs on the leading edge where EUV lives, there doesn't seem to be room for a second tool vendor to compete with ASML, so this may be the new normal.

Well, yeah, to your point... is Samsung supposed to outbid TSMC for ASML machines with those prepaid QC orders? I mean, if we are just running into supply chain issue with growing fabs (seemingly yes) then everyone throwing more money at the wall doesn't do much unless it grows the production capacity of the component creating the issue.

And if this is the new normal, it sucks for everyone. Eventually it will suck for Apple too when they reach a point where they are either being A) having costs go up significantly due to competitive bids on future silicon or B) having to delay a new generation of anything due to there being some hiccup in production which is tightly constrained. Or both I guess. The older risk mitigation factors of using multiple suppliers doesn't work so well when their is one game in town.

It's a bummer.

I am rooting for Samsung AND Intel to really up their game because I want shiny things and I don't want to pay $2k for a GPU, for example.

About the article itself - I was kind of shocked at how poor the Samsung chip did with regards to perf/w compared to such a similar QC implementation.

And the graphics performance all being on the back of unsustainable power usage for both chips? It seems like a mess.

Also skipped all those graphs and just read the text.
 
Last edited:
  • Like
Reactions: Tlh97

gdansk

Golden Member
Feb 8, 2011
1,979
2,355
136
If 5nm LPE is holding them back why would they bother?
That extra die space being available for the GPU is perhaps the "achievement" of 5nm LPE.
 
Last edited:

Doug S

Platinum Member
Feb 8, 2020
2,202
3,405
136
And if this is the new normal, it sucks for everyone. Eventually it will suck for Apple too when they reach a point where they are either being A) having costs go up significantly due to competitive bids on future silicon or B) having to delay a new generation of anything due to there being some hiccup in production which is tightly constrained. Or both I guess. The older risk mitigation factors of using multiple suppliers doesn't work so well when their is one game in town.

Apple can't be "outbid", they are paying years in advance. No one else is willing to risk that, when they don't know the demand for their products. But yes, they are vulnerable if TSMC runs into problems, but they are apparently working with TSMC on developing new nodes now so they will have plenty of notice so it won't catch them with their pants down.

The end of Moore's Law won't be reached from physical limits, but from economic limits. Eventually there will be so few customers willing to go to a new node that it won't pay for TSMC (or whoever is left alone at the leading edge if it isn't them) to develop it. Every new node has fewer customers because it requires more wafer runs per design to make sense.
 
  • Like
Reactions: Tlh97 and moinmoin

blckgrffn

Diamond Member
May 1, 2003
9,110
3,029
136
www.teamjuchems.com
Apple can't be "outbid", they are paying years in advance. No one else is willing to risk that, when they don't know the demand for their products. But yes, they are vulnerable if TSMC runs into problems, but they are apparently working with TSMC on developing new nodes now so they will have plenty of notice so it won't catch them with their pants down.

The end of Moore's Law won't be reached from physical limits, but from economic limits. Eventually there will be so few customers willing to go to a new node that it won't pay for TSMC (or whoever is left alone at the leading edge if it isn't them) to develop it. Every new node has fewer customers because it requires more wafer runs per design to make sense.

Well, someone else who thinks it's absolutely vital (life and death) could offer to also invest years ahead. There might already be other agreements coming together - and even if that exists it will drive competition and the price for Apple to keep their position will go up. We could spitball a list of fairly massive customers who might be willing to put their names down.

Haha, you triggered me buy using some absolute terms there. I'll refuse to concede that there is absolutely no other possibilities in existence :D

Obviously Apple has the cash to do what they want but other companies are able to see the writing on the wall - get in early or spend ever longer times on trailing edge nodes.
 

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
Well what's the alternative?

TSMC really ought to have served all the customers that they had served in the recent past - at least among those who were still willing to continue to effectively fund continuing node advancements. Perhaps Qualcomm just didn't want to pay to play at the same level Apple did, which is why Apple wound up with the node all to themselves. Maybe Samsung offered them a sweetheart deal, or maybe Qualcomm is just trying to throw money at another fab to try to bring about some parity. That's a lot of speculation, though.

Qualcomm showed up for N7 and N7P, so it's hard to believe that they weren't willing to commit to N5 had they been given the chance.
 
  • Like
Reactions: Tlh97

moinmoin

Diamond Member
Jun 1, 2017
4,934
7,619
136
I get the feeling many of you think of TSMC as some store with ready-made products and customers fighting each other to get in and get some of the much sought after products.

That's not how it goes, TSMC's store is empty and customers essentially order the products well in advance, Apple even before TSMC goes ahead building the production capacity, everybody else once that has settled somewhat. This is multiple years before the actual finished products can be sold to consumers, and even with all preparations done the lead time for new orders (not counting the time for masks, validation etc.) is more than half a year. There's not so much an inability to serve customers but a backlog of orders that is being fulfilled as soon as capability is available, while said capability is being expanded in multiple phases at the same time.

If Qualcomm doesn't use TSMC this has been a conscious decision by Qualcomm, for whatever reason. E.g. MediaTek, widely considered a lesser competitor, had no such issues.
 
Mar 11, 2004
23,031
5,495
146
Yeah this is a move Qualcomm made for some reason, very likely before we knew the lay of the land (Samsung's 7 and 5 nm not being great). My guess is Samsung offered some very good deal, which is why Nvidia went that route (and they likely got burned as well). They might have been anticipating a demand crunch (its been brewing for some time) as well. And they might have believed that Samsung's 7 and 5 nm would be doing well (Samsung made a big investments on it and maybe they thought them being EUV would make a lot of difference). I'd imagine it was a pricing thing, possibly coupled with Samsung guaranteeing buying a certain amount of chips for their phones.
 

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
If Qualcomm doesn't use TSMC this has been a conscious decision by Qualcomm, for whatever reason. E.g. MediaTek, widely considered a lesser competitor, had no such issues.

I still have a hard time Qualcomm could be that crazy, especially when they hedged with the Snapdragon 870.
 
  • Like
Reactions: Tlh97

Ajay

Lifer
Jan 8, 2001
15,332
7,792
136
There is just not enough Fab production capacity on cutting edge nodes for TSMC to service all customers in a timely matter. Apple, as the company that pushed TSMC to the fore front with massive upfront orders and the willingness to purchase risk wafer (with many dice being tossed into the bin) has really subsidized TSMC's advanced node development more than any other company. AMD has a great relationship with TSMC and orders very large numbers of wafers - plus gives TSMC the PR advantage of producing high clocked, high performance CPUs. No doubt NV blew it's relationship to some extent by moving over to Samsung. I think QC got a good deal with Samsung (they were offing FREE mask sets for a while) and, as they compete with Apple only indirectly, having a lower price is likely an asset. QC also demonstrated their engineering prowess by extracting much better perf/watt numbers out of the same node as Samsung. It's really too bad that Samsung just can't get closer to TSMC, but competing on prices, not performance, limits their ability to attract the best talent and invest early in the best equipment; at least on the logic side.
 
Last edited:

Doug S

Platinum Member
Feb 8, 2020
2,202
3,405
136
TSMC really ought to have served all the customers that they had served in the recent past - at least among those who were still willing to continue to effectively fund continuing node advancements. Perhaps Qualcomm just didn't want to pay to play at the same level Apple did, which is why Apple wound up with the node all to themselves. Maybe Samsung offered them a sweetheart deal, or maybe Qualcomm is just trying to throw money at another fab to try to bring about some parity. That's a lot of speculation, though.

There's a good chance they got a better deal from Samsung. 1) because TSMC has NO incentive to offer any discount at all when they are full/beyond capacity at N7 let alone N5. 2) because Samsung has not been able to fill their leading edge logic (it is a different story for DRAM/NAND of course) fabs for years. That may no longer be true with the EUV nodes though given the ASML limitation, but until I hear some evidence that Samsung's fabs are even close to full at 5nm I'll assume there is slack capacity like they've always had.
 
  • Like
Reactions: Tlh97

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
Am I the only one who finds Andrei's graphs ridiculously hard to grok? He has loads of great information, but the presentation sucks. Graph two different variables in opposing directions, for 10 different devices with hard to track colour coding.

EDIT: Specifically the graphs like this:

SPECint06.png

Maybe they're a bit tough at first, but once you get it they're really good and make as much use of space as possible. Bars on either side that are farthest to the left (least power, most performance) are best and colors and shades of them represent different manufacturers or product families. In this case anything shaded orange is a Qualcomm SoC.

Trying to present that any other way would be a massive page of separate graphs or having to scroll through some JavaScript carrousel which makes direct comparison between different benchmark parts more difficult. I've never seen something like that used before, but once I got it, I though it was absolutely brilliant. About the only thing that could make it better would be a mouse over to show percentage difference between the bars, but eyeballing them will get you close enough.
 

moinmoin

Diamond Member
Jun 1, 2017
4,934
7,619
136
Andrei did write that with 5LPE Samsung only caught up with TSMC's N7 efficiency wise, and that only at lower frequencies. Except possibly from a price point Samsung doesn't appear to be competitive with TSMC's latest node.
 
  • Like
Reactions: Tlh97

blckgrffn

Diamond Member
May 1, 2003
9,110
3,029
136
www.teamjuchems.com
Tangent: Anandtech shared a new announcement of new QC 5G on "4nm" Samsung.

Is that just another 7nm++++ variant for Samsung or something interesting in that we could see 4nm mobile CPUs that are different at all from the current 5nm parts?
 

Hitman928

Diamond Member
Apr 15, 2012
5,180
7,631
136
Tangent: Anandtech shared a new announcement of new QC 5G on "4nm" Samsung.

Is that just another 7nm++++ variant for Samsung or something interesting in that we could see 4nm mobile CPUs that are different at all from the current 5nm parts?

It's an evolution of 7LPP. The next distinct (if that's what you want to call it) node step is 3GAE.
 
  • Like
Reactions: Tlh97 and blckgrffn

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
There's a good chance they got a better deal from Samsung.

That was also my suspicion . . . I just have to wonder: did Qualcomm know it would end up this way? I could see dual-sourcing Snapdragon 888 at TSMC and Samsung to take advantage of the deals and use the "bad" 888 chips in certain devices where Qualcomm might feel they can get away with it. But going entirely with Samsung? That's ballsy or just plain stupid.
 

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
Qualcomm may not have had much of a choice in the matter about going with Samsung. I've just penned a one act play about it.

TSMC: Hello Apple, old buddy, how are things going?

Apple: Really good. These 7nm chips have certainly been amazing.

TSMC: That's great. We're just getting 5nm ready to go, and it's going to be even more amazing.

Apple: That's certainly sounds enticing.

TSMC: It's definitely going to be a hot commodity. We could probably get you down for 10,000 wafers no problem.

Apple: Actually we want all of it.

TSMC: All of it? I mean that's great, but it is a new node and it is a bit pricey.

Apple: Yeah, whatever. Just take this ridiculously large sack of money and just give us all of the wafers you can make.

TSMC: Pleasure doing business with you!

Qualcomm: Hey TSMC, haven't talked to you in a while. I've been hearing some good things about this 5nm stuff you're working on.

TSMC: Sorry, we're busy counting all of this money right now. Begone peasant!