Two Questions for Intel CEO Brian Krzanich

[Daniel Nenni]

View attachment 17373For those of you who are frustrated by the questions asked by analsysts during conference calls here is a good one. Intel hosted a "virtual" online stockholder meeting on May 19 and there were two very telling questions asked by a Motley Fool. Considering that Mr Fool has zero semiconductor experience or education they were good questions. His analysis of the answers however were absurd. Just more media regurgitation of press releases and power point slides.

Manufacturing Technology

Fool:
Your competition have said that they will transition to 10-nanometer manufacturing by the end of the year, earlier than Intel will. Can investors have confidence that Intel has a clear manufacturing lead?

Krzanich:
My commitment to you is that yes, you can, as an investor, count on Intel having a clear manufacturing lead. You have to remember these measurements don't necessarily mean all the same thing and how people are measuring the 10-nanometers versus 14-nanometers is different. You really need to take a look at the performance of the devices and what's being delivered. And, so, we believe our 14-nanometer technology will still outperform any of the other technologies that'll be introduced in the next several years. Our 10-nanometer technology will again make another large leap in performance and be even further ahead than we think the 14-nanometer will be.

Competition in Servers

Fool:
On competition in servers, does Intel believe that it can continue to maintain clear product leadership over Advanced Micro Devices, even as it rolls out its new Zen-based PC and server products?

Krzanich:
I'm not going to comment on just one company, I'm just going to take a look at the industry in general and say: If we look at all of our competition out there, yes we believe that our sixth generation Xeon E7 family of products will continue to be a leadership product, far and away leaders over all of our competition. And, you'll continue to see a series of products that are announced this year that will continue to drive that performance. So, yes, we're very comfortable with our server products and will continue to have leadership. And, you'll continue to see a series of products that are announced this year that will continue to drive that performance. So, yes, we're very comfortable with our server products and will continue to have leadership.

I'm on my way to Taiwan today so I will comment later but I really do think this deserves discussion....

Especially the first question, does Intel or anyone else for the matter believe that Intel still has a manufacturing lead? If so, what does "manufacturing lead" mean?

Cheers,

D.A.N.

 

[hist78]

My commitment to you is that yes, you can, as an investor, count on Intel having a clear manufacturing lead.

Manufacturing lead didn't propel Intel into mobile product dominance. Because there are many factors need to be considered in a customer's mind. Manufacturing lead doesn't guarantee the final outcome of a processor.

Using the new Google's TPU as an example. It's almost for sure Google is not using Intel's x86 architecture in their TPU. So the big question to Intel is: if Google is not buying your product and Google is not using your architecture, does that mean your "manufacturing lead" is much less important and much less relevant?

 

[Jozo035]

You really need to take a look at the performance of the devices and what's being delivered.

This is what i am saying and i absolutely agree with this. But:

FPGA: 14nm was not delivered.
GPU/accelerators: 14nm KNL - ~3 TFlops vs 16nm Pascal P100 - 5.3 TFlops.
mobile SOC's: Intel admitted defeat.
desktop: AMD still with 32nm process, so first point for Intel.
HPC: ARM, OpenPOWER, Sonoma... still in development, second point for Intel.
Mainframes: they are trying so hard but still can't compete even 22/20nm IBM and Sun devices.

2/6 points for Intel, when we don't count ASIC, IoT, automotive... And it can go worse with introduction of 14nm AMD processors.

 

[astilo]

This is what i am saying and i absolutely agree with this. But:

FPGA: 14nm was not delivered.
GPU/accelerators: 14nm KNL - ~3 TFlops vs 16nm Pascal P100 - 5.3 TFlops.
mobile SOC's: Intel admitted defeat.
desktop: AMD still with 32nm process, so first point for Intel.
HPC: ARM, OpenPOWER, Sonoma... still in development, second point for Intel.
Mainframes: they are trying so hard but still can't compete even 22/20nm IBM and Sun devices.

2/6 points for Intel, when we don't count ASIC, IoT, automotive... And it can go worse with introduction of 14nm AMD processors.

Still a point for Intel, but AMD desktops are at 28nm since 2014. In 2017 both Intel and AMD will be at 14nm, so time will tell.

 

[benb]

This forum doesn't find much to like when BK speaks--we've plowed this ground a lot. The optics of Intel slowing development are not good.

In terms of competitive pressure, differentiation and low cost: Intel cost is hard to analyze, but their margins are not--Intel margins are high. That is an advantage. Their high margins will attract competition. The same thing happened with TSMC, which has equally high margins; this attracted Samsung and GF to compete with them. Intel has fewer competitors than TSMC.

Intel differentiation is also an advantage. Where TSMC, Samsung and GF all make the same chips, Intel has a moat--they own IP on the chips they make. This is a huge advantage and puts Intel in a different category than every other chip company I can think of. Most of the foundry/fabless industry is engaged in making low-differentiation, yet very expensive, ARM chips; it is hard to build a sustainable business on this basis. Qualcomm in particular is constantly at risk of becoming a victim of commoditization. Intel doesn't have that problem.

In the long run though, staying ahead of commoditization is difficult. AMD for example has fallen victim to it. It's possible Intel is underestimating how easily they could become like AMD.

 

[Daniel Nenni]

The problem I have is that the definition of "Process Lead" is a moving target. Transistor speed? Density? Power? Cost? Or some combination thereof? And where exactly is BK getting this information?

How about we look at which processes are the best?

In head-to-head comparisons, Apple and QCOM put out the "best" SoCs, much better than Intel, on comparable foundry processes. Unfortunately, after spending $10B+ on mobile, Intel is out of the SoC business but we can still look at their modems when they are on an Intel 14nm process.

It will also be interesting to see how the AMD 14nm Zen compares to the latest Intel CPU offering in regards to process technologies. I can tell you one thing, the Samsung 14nm wafers are MUCH less expensive than Intel 14nm.

While I do agree that process names are misleading, one thing I can tell BK today is that TSMC 16nm FFC is a "better" process than Intel 14nm and TSMC 10nm will be even "better", absolutely.

 

[Staf_Verhaegen]

The problem I have is that the definition of "Process Lead" is a moving target. Transistor speed? Density? Power? Cost? Or some combination thereof? And where exactly is BK getting this information?

What is a fact is that Intel is the only company that has tried to keep the node number somewhat related to Moore's law scaling. The other's have given that up the moment they called 20nm+Finfet 16/14nm.

 

[astilo]

What is a fact is that Intel is the only company that has tried to keep the node number somewhat related to Moore's law scaling. The other's have given that up the moment they called 20nm+Finfet 16/14nm.

I agree that 1nm is not the same beast for foundries and IDMs, but I disagree about Intel being the only company trying to keep the node number correlated somehow to the technology node critical dimensions. There is nothing, for instance, with a 22nm feature on such a node.
And no new node is bigger or less dense than the previous one, and this is true for both IDMs and foundries.
Moore's law is no law at all if you ask me.

 

[Jozo035]

https://www.semiwiki.com/forum/content/3099-tsmc-responds-intels-14nm-density-claim.html

 

[Daniel Nenni]

What is a fact is that Intel is the only company that has tried to keep the node number somewhat related to Moore's law scaling. The other's have given that up the moment they called 20nm+Finfet 16/14nm.

I disagree. It was FinFETs that really gave marketing people creative process naming license. Is the Intel 22nm process really a 22nm process by the industry definition? How can you accurately compare the minimum length of a planar transistor to that of a FinFET?

View attachment 17387

I'm okay with marketing people having their fun but I'm not okay with the CEO of a semiconductor company with 30 years of manufacturing experience saying that one FinFET process node name is more "accurate" than another.

It actually started before FinFETS with people taking creative license with the term minimum gate length:

View attachment 17386


Bottom line: You can no longer say one process node name is technically more accurate than another, unless of course you are a marketing person.

 

[benb]

This isn't that difficult of an issue--in fact it's trivial to determine whether a shrink is truly a shrink. Chips are square pieces of silicon with a certain number of transistors. A shrink from node A to node B, must be a 60+% reduction in area for the same number of transistors. Or else it's a fraud. And fraud will get you sent to jail. The law allows for puffery, not fraud. You can call a 100 HP engine 104 HP, but not 200 HP.

If you call your 60+% shrink "14nm" and someone else calls it "22nm", what's the harm? Intel has three systems I can think of--"1272", "Fifth Generation Core", Core i5-5xxx, none of them referencing "14nm". So, clearly, the shrinks the thing, not the name.

 

[Jozo035]

[table] style="width: 693px"
|-
| style="height: 20px; width: 64px" | Node name
| style="width: 99px" | Transistors (M)
| style="width: 123px" | Area (mm2)
| style="width: 321px" | Area of 1M of average transistors (Area/Transistors)
| style="width: 87px" | Area scaling (r/r+1)
|-
| align="right" style="height: 20px" | 14
| align="right" | 1750
| align="right" | 122
| align="right" | 0,069714286
| align="right" | 0,551412429
|-
| align="right" style="height: 20px" | 22
| align="right" | 1400
| align="right" | 177
| align="right" | 0,126428571
| align="right" | 0,678968254
|-
| align="right" style="height: 20px" | 32
| align="right" | 1160
| align="right" | 216
| align="right" | 0,186206897
| align="right" | 0,517556051
|-
| align="right" style="height: 20px" | 45
| align="right" | 731
| align="right" | 263
| align="right" | 0,359781122
| align="right" | 0,732142003
|-
| align="right" style="height: 20px" | 60
| align="right" | 291
| align="right" | 143
| align="right" | 0,491408935
| align="right" | 0,500343643
|-
| align="right" style="height: 20px" | 90
| align="right" | 112
| align="right" | 110
| align="right" | 0,982142857
| align="right" | 0,545634921
|-
| align="right" style="height: 20px" | 130
| align="right" | 45
| align="right" | 81
| align="right" | 1,8
|
|-
[/table]




This is comparison of transistor count's of Intel's quad core processors (since 45 nm) with integrated GPU (since 32nm, but it seem that GPU does not really matter). Numbers in last column means, that for example 14nm needs 55,14% of area of 22nm to maintain same count of transistors.

A shrink from node A to node B, must be a 60+% reduction in area for the same number of transistors.

Does that mean, that new chip with similar count of transistor should have less than 40% of area of old chip (100% - 60% reduction).
Or chip with same size should have 1/(1-0,6)=2,5 times more transistors?

Because if yes, then Intel is not even close to that. Especially with 22 and 45nm.

 

[benb]

Jozo, this table is interesting. It makes me wonder if Intel has any transistor-level redundancy which they conceal from the official transistor count. And whether that redundancy varies from node to node based on their business plans.

I imagine Intel's position in the server market is in part set by the high reliability of their transistors--Enterprise grade. The only way to ensure that level of reliability is redundancy.

 

[ippisl]

To me one interesting question regarding Intel's competitiveness: How come they still own the chromebook cpu market - when Samsung should have all the motivation to compete and sell the full package?

Do they lose money on that processor line ? will they stop making it some time in the future ?

 

[tty2]

one thing I can tell BK today is that TSMC 16nm FFC is a "better" process than Intel 14nm and TSMC 10nm will be even "better", absolutely.

Based on what? How bout we wait for the pudding?

 

[Portland]

Intel chose the wrong guy.