[WSJ] Samsung, Nvidia Shy Away from Server Chip Battle

[NTMBK]

I can imagine the biggest players using this product as a stepping stone to develop their own SoCs with hard logic that implements the same functions more efficiently than an FPGA

Hard logic has a) very long development cycles b) no flexibility. Modern web giants thrive by rapidly iterating and improving on their software, not locking it down for multi-year cycles.

 

[Exophase]

Hard logic has a) very long development cycles b) no flexibility. Modern web giants thrive by rapidly iterating and improving on their software, not locking it down for multi-year cycles.

The use of accelerators in FPGAs is totally uncharted, you can't directly apply their software development cycles to what you expect to happen there. After a couple years of doing this it could end up being obvious what core functions would benefit from hard accelerators and not need to be constantly iterated. We're talking about a very specific and much simpler tiny subset of their current functionality.

 

[Arachnotronic]

But Intel's "massive growth" phase is over- the PC market is stagnant, and they are failing in the phone and tablet markets. Where will the new volume come from to fill those leading edge fabs?

Intel is operating at very high utilization levels now (>80%) as you can see in the company's updated gross margin outlook for the quarter (64% -- this is near 2010 highs).

At any rate, Intel sells, what? $49-$50 billion worth of logic semis? Who else has that kind of revenue?

The next closest player, TSMC, is still playing in the $20 billion/year revenue sandbox and -- if Samsung's foundry business takes off and/or Intel gains traction in mobile with its own products -- that could come down significantly.

No, Intel has a big fat high margin DCG business protected by incumbency and R&D might, and with respect to the gigantic PC client group business we're now seeing signs of stabilization.

Intel has the volume/scale here, and if it can just get its mobile act together (I realize this is a big "if" given how absolutely terrible its execution has been thus far), then it will be in a pretty good position.

 

[Homeles]

Intel is operating at very high utilization levels now (>80%) as you can see in the company's updated gross margin outlook for the quarter (64% -- this is near 2010 highs).

At any rate, Intel sells, what? $49-$50 billion worth of logic semis? Who else has that kind of revenue?

The next closest player, TSMC, is still playing in the $20 billion/year revenue sandbox and -- if Samsung's foundry business takes off and/or Intel gains traction in mobile with its own products -- that could come down significantly.

No, Intel has a big fat high margin DCG business protected by incumbency and R&D might, and with respect to the gigantic PC client group business we're now seeing signs of stabilization.

Intel has the volume/scale here, and if it can just get its mobile act together (I realize this is a big "if" given how absolutely terrible its execution has been thus far), then it will be in a pretty good position.

Last I checked, Samsung was around $30 billion/yr.

 

[Roland00Address]

The question is not whether Arm and other architectures can compete but whether there is enough return on investment, it is easy and cheap to make an arm chip. The only question is whether its worth going up against intel and is the ROI worth it. Intel has so many structural advantages that if they choose to they will always be cheaper than you (you being an arm maker).

The only other way to make money via arm is be far easier to integrate X where X is something novel. In this case it is not the cost that matters but the simplicity and time to market issues.

But Intel's "massive growth" phase is over- the PC market is stagnant, and they are failing in the phone and tablet markets. Where will the new volume come from to fill those leading edge fabs?

Google I/O started just yesterday and the big focus was not computers but instead everything else. We are talking about tablets, phones, cars, tvs, watches, sensors, etc. Electronics are not going away they are just getting smaller and cheaper to make and due to this they will be more plentiful in the world.

The days of small volume but expensive chips costing hundreds of dollars is over, the new future is high volume small chips, aka lots of small chips that cost 20 dollars or less and eventually have costs measured in cents not dollars.

 

[Arachnotronic]

Last I checked, Samsung was around $30 billion/yr.

Almost all of that is DRAM/NAND. Very little logic. In fact, I'd bet that the vast majority of Samsung's logic chip production is Apple A-series processors.

 

[Homeles]

Almost all of that is DRAM/NAND. Very little logic. In fact, I'd bet that the vast majority of Samsung's logic chip production is Apple A-series processors.

Ah, okay. I goofed on that one.

Google I/O started just yesterday and the big focus was not computers but instead everything else. We are talking about tablets, phones, cars, tvs, watches, sensors, etc. Electronics are not going away they are just getting smaller and cheaper to make and due to this they will be more plentiful in the world.

The days of small volume but expensive chips costing hundreds of dollars is over, the new future is high volume small chips, aka lots of small chips that cost 20 dollars or less and eventually have costs measured in cents not dollars.

Definitely. Bay Trail was rather lackluster when it came to the bill of materials. I mean, is it just me, or is there a ton of wasted space on die?

This is where Intel could really use high levels of design automation. There's about, what, 10-20% of the die that's effectively nothing but empty space? Compare to AMD's Kaveri/Beema, and you'll see that even AMD has better utilization of die space. Compare to ARM SoCs, which are jam-packed.

Bay Trail is not terribly cost competitive with Haswell Celerons.

However, to get to what you were saying, this is going to change, even in the short term. Cherry Trail and the like are going to be better on this front. The selling price of Cherry Trail-D is obviously going to be highly subject to competition, but the cost to manufacture will be considerably lower. Intel will be catching up to their competition in when it comes to cost, which is their only real weak point (although a large one) right now.

Combined with the "More-than-Moore" density increase that 14nm is bringing (~2.2x), the cost implications could be considerable. Even Broadwell is supposed to improve on cost-optimization, irrespective of the new process, which is already very cost-focused.

And it's only going to continue from there. ARM and AMD might hit a lull with 20nm and 16/14nm, but 10nm is going to be a very big improvement on the cost front for Intel and everyone else as well. I actually wouldn't be surprised if these savings end up being passed down to x86 laptops and desktops. I'd be willing to guess that the Cannonlake equivalent to the i5-4690K will sell for less than $200, and the i7-4790K equivalent will be less than $300, assuming these designs stay quad cores. If the price doesn't drop, I'd expect very large changes to be made to justify the higher price (e.g., an integrated PCH on the whole lineup may keep pricing flat, but reduce platform cost).

However, these industry-wide cost reductions also going to squeeze players out of the market. Hopefully semiconductor growth is high enough to mitigate this.

People may be disillusioned with the current rate of progress, but that's going to change within the next 2 years.

 

[tarlinian]

Definitely. Bay Trail was rather lackluster when it came to the bill of materials. I mean, is it just me, or is there a ton of wasted space on die?

That's not what bill of materials refers to. The chip itself is just fine in terms of cost. The problem is that it needs a lot more supporting parts than competing ARM SoCs to make a functioning tablet, thereby requiring subsidies to make it work. And the issue here isn't wasted space either. Intel has much much less of an advantage in internal expertise is analog/mixed signal work. It also hasn't been really valuable in a long time so they haven't focused on it. (e.g., there's no point to integrating a $10 device controller when the chip sells for $500.) As average cost per chip comes down, integration matters more.

 

[Homeles]

That's not what bill of materials refers to. The chip itself is just fine in terms of cost. The problem is that it needs a lot more supporting parts than competing ARM SoCs to make a functioning tablet, thereby requiring subsidies to make it work. And the issue here isn't wasted space either. Intel has much much less of an advantage in internal expertise is analog/mixed signal work. It also hasn't been really valuable in a long time so they haven't focused on it. (e.g., there's no point to integrating a $10 device controller when the chip sells for $500.) As average cost per chip comes down, integration matters more.

I understand what BoM is. It seems you have misunderstood what I was saying.

 

[itsmydamnation]

Almost all of that is DRAM/NAND. Very little logic. In fact, I'd bet that the vast majority of Samsung's logic chip production is Apple A-series processors.

samsung is shifting more and more phones/tablets to exynos, note , note 10.1, S, tab pro, nexus etc .

last time i looked the note 3 was moving 5 mill a month , S5 10 mill a month , the other products are harder to get sales figures on but combined they would be ~3 mill or so a month.

Samsungs problem at the moment is low mid end phone competition where they don't make there own chips.

So i wouldn't go around assuming that the only sizable logic samsung makes is for apple.

off topic put personally i think sony is schooling everyone (including apple) in product integration and physical design at the moment so it will be interesting to see how that plays out of the next couple of generations.

 

[Arachnotronic]

samsung is shifting more and more phones/tablets to exynos, note , note 10.1, S, tab pro, nexus etc .

last time i looked the note 3 was moving 5 mill a month , S5 10 mill a month , the other products are harder to get sales figures on but combined they would be ~3 mill or so a month.

Samsungs problem at the moment is low mid end phone competition where they don't make there own chips.

So i wouldn't go around assuming that the only sizable logic samsung makes is for apple.

off topic put personally i think sony is schooling everyone (including apple) in product integration and physical design at the moment so it will be interesting to see how that plays out of the next couple of generations.

Erm...Galaxy Note 3 in NA was S800, S5 in NA is S801, LTE versions of all of Samsung's tablets are S800, and even the low end Galaxy Tab 4 is Qualcomm S400 across the board.

Also, 10 million/month for S5? Yeah, maybe for the first month, but it falls off dramatically from there. I would be surprised if Samsung could muster more than 3-5 million/month after maybe the first month where it satisfies pent up demand + stuffs the channel.

So Samsung uses Exynos in Wi-Fi only tablets at the high end and non-LTE versions of its handsets. It's just not that much volume for the R&D that Exynos is probably sucking up.

 

[itsmydamnation]

the world is a lot bigger then just NA, outside of NA its almost always exynos.

Also, 10 million/month for S5? Yeah, maybe for the first month, but it falls off dramatically from there.

anything to back up the assumption ? while there is always a decline over time, its never been "dramatic" over just a single month. But thats why they have staggered product launches.

 

[tarlinian]

I understand what BoM is. It seems you have misunderstood what I was saying.

What does BOM cost have to do with die space usage then? Inefficient use of space affects die cost, not how much the other components of the system cost.

Unless you were obliquely trying to imply that the wasted space could have been used for functionality which currently requires discrete components and increases BoM cost, your post doesn't make sense. And if you were trying to imply that, I think it's pretty well understood that missing functionality in Bay Trail is due to both design decisions early in its cycle that were intended more for PCs (how many ARM SoCs have SATA controllers?) and lacking a breadth of design expertise and a lack of external IP suitable for their internal fabs, not because of empty space left on the chip by a lack of design automation. These factors should both improve as Intel focuses more on this area.

 

[AtenRa]

Ah, okay. I goofed on that one.

Definitely. Bay Trail was rather lackluster when it came to the bill of materials. I mean, is it just me, or is there a ton of wasted space on die?

This is where Intel could really use high levels of design automation. There's about, what, 10-20% of the die that's effectively nothing but empty space? Compare to AMD's Kaveri/Beema, and you'll see that even AMD has better utilization of die space. Compare to ARM SoCs, which are jam-packed.

Bay Trail is not terribly cost competitive with Haswell Celerons.

However, to get to what you were saying, this is going to change, even in the short term. Cherry Trail and the like are going to be better on this front. The selling price of Cherry Trail-D is obviously going to be highly subject to competition, but the cost to manufacture will be considerably lower. Intel will be catching up to their competition in when it comes to cost, which is their only real weak point (although a large one) right now.

Combined with the "More-than-Moore" density increase that 14nm is bringing (~2.2x), the cost implications could be considerable. Even Broadwell is supposed to improve on cost-optimization, irrespective of the new process, which is already very cost-focused.

And it's only going to continue from there. ARM and AMD might hit a lull with 20nm and 16/14nm, but 10nm is going to be a very big improvement on the cost front for Intel and everyone else as well. I actually wouldn't be surprised if these savings end up being passed down to x86 laptops and desktops. I'd be willing to guess that the Cannonlake equivalent to the i5-4690K will sell for less than $200, and the i7-4790K equivalent will be less than $300, assuming these designs stay quad cores. If the price doesn't drop, I'd expect very large changes to be made to justify the higher price (e.g., an integrated PCH on the whole lineup may keep pricing flat, but reduce platform cost).

However, these industry-wide cost reductions also going to squeeze players out of the market. Hopefully semiconductor growth is high enough to mitigate this.

People may be disillusioned with the current rate of progress, but that's going to change within the next 2 years.

I will just leave those here, i really dont think that 14nm will bring cost down unless you seriously shrink your die size. By doing that you compromise on performance (less xtors less performance smaller die). So, either you compromise on performance by reducing die size or you increase sales by trying to get a hold to more markets. Intel is trying both of them.
10nm will be even worst.

 

[Arachnotronic]

the world is a lot bigger then just NA, outside of NA its almost always exynos.

That's actually not true. Outside of NA, Samsung uses a lot of Qualcomm chips, some Marvell here and there, and recently (although not going forward) Broadcom.

ImgTec's CEO estimated that only 10-20% of Samsung's apps processors were Exynos, with the rest sourced from external vendors.

Samsung just doesn't have the modem capability that Qualcomm and some of the others have, which is why it's just easier for Samsung to buy integrated off-the-shelf parts than to try to roll its own.

Exynos is really a mess, IMO, and the sooner it can be put out of its misery the better. Samsung would be perfect as a foundry for all of its apps processor vendors which is where I think their semi business ultimately goes.

 

[itsmydamnation]

ImgTec's CEO estimated that only 10-20% of Samsung's apps processors were Exynos, with the rest sourced from external vendors.

i dont disagree with that, i think people underestimate how many low end phones samsung sells and none of those are exynos. It might change this generation but the S4 note II era most of the 1st world outside of the USA got a 3g oct core based exynos. with LTE deployments now being much larger i don't think they can get away with it from a marketing point of view ( actual data useage on the other hand......)

IDC estimates 314million samsung phones in 2013 shipped. 10-20% is 31 to 60 million chips, that is still 3,631,032,000mm2 to 7,262,064,000 mm2 of SOC a year. that isn't insignificant, now im late for work D:

 

[ShintaiDK]

the world is a lot bigger then just NA, outside of NA its almost always exynos.

Thats not true. The international versions are Qualcomm.

For the S4 5 of the 6 versions didnt use Samsung either. And the same applies for most of their products.

 

[DrMrLordX]

While I have no doubt that "some" of Intel's server growth is coming from 'cloud companies', how do we know that "most" of the server growth is coming from that segment?

Does Intel do revenue breakouts by business division? (honest question, I wasn't aware Intel went into that granularity with their revenue streams) Is this a new thing with BK becoming CEO?

I did some (admittedly cursory) research on actual server revenues over the last . . . seven years or so. If you look at numbers from the "box counters" you'll see global server revenues kicking around at about $50 billion per year; however, many critics of this data contend that ODM shipments are grossly underrepresented in those findings.

The problem is that some firms place orders with ODMs that essentially go under the radar.

Based on what I've seen, Intel isn't overly-enthusiastic about providing specific data to the public regarding server CPU shipments.

 

[witeken]

I will just leave those here, i really dont think that 14nm will bring cost down unless you seriously shrink your die size. By doing that you compromise on performance (less xtors less performance smaller die). So, either you compromise on performance by reducing die size or you increase sales by trying to get a hold to more markets. Intel is trying both of them.
10nm will be even worst.

http://oi62.tinypic.com/2yltg6h.jpg

http://oi57.tinypic.com/166yal2.jpg

http://oi59.tinypic.com/34il5yr.jpg

Nope.

 

[NTMBK]

Nope.

That graph is cost per transistor. Does not take initial capital costs (i.e. fabs) into account.

 

[AtenRa]

Nope.

We've been through this before,

http://forums.anandtech.com/showpost.php?p=36044325&postcount=56

 

[kimmel]

That graph is cost per transistor. Does not take initial capital costs (i.e. fabs) into account.

It wouldn't make business sense to continue on shrinking if it didn't. It would be an absolutely pointless graph. When you get wafer pricing from TSMC/Intel/GloFo of course fab costs factor in to the price they charge you.

 

[mrmt]

That graph is cost per transistor. Does not take initial capital costs (i.e. fabs) into account.

CAPEX only matter if you don't get your targeted volumes. I can see why Globalfoundries might be shy of 10nm and 7nm, but not Intel and TSMC. The rest of the foundries won't even try.

 

[ShintaiDK]

There is an old list from IHS. While not entirely accurate, it does show the tendency. GloFo for example essentially gave up on 20nm with FF and licensed it from Samsung. And STM is out. I wonder if there will be more than 1 company left for 7 and 5nm.

 

[witeken]

Samsung and Intel, but I'm not so sure about TSMC. It depends on how much market share Intel will grab of the internet of things, tablet and smartphone markets, and Intel won't stop at 30%. TSMC basically gives up at 20nm.