EETimes: TSMC starts FinFETs in 2013, tries EUV at 10 nm

[Piroko]

But it *is* billions of dollars.

TSMC is spending 1.4 billion dollars in R&D alone in 2012, up from 1.1 billion in 2010. But this number doesn't tell the whole story. As node takes 4-5 years in the R&D pipeline, it's safe to say that the next future node will cost at least 3x the amount of money of the current R&D expenditure. 20nm Should cost TSMC some 3-3.5 billion, 16nm will cost some 4 billion.

Those sums only add up if TSMC isn't spending anything at all on unrelated R&D and only throw out a single process every 2.5 - 3 years.

But R&D alone doesn't mean anything, the retooling of the factory is also important, and here comes the Rock's law. "The cost of a semiconductor chip fabrication plant doubles every four years". Meaning your CAPEX double every four years. This is what crushed AMD in 2009 and this is what keep SMIC, UMC, IBM and others from competing with TSMC, Intel and Samsung for the bleeding edge node. Even if they had the R&D, they cannot afford the fabs.

So in order to research 20nm *and* build a single factory with significant capacity you need some 12 billion dollars, spread around 4-5 years. TSMC current CAPEX is 9 billion dollars, or more than the entire GLF investment alone.

Underlining my point, CAPEX isn't R&D and should not be thrown into the same pot. But it is a much more valuable number to compare.

Do you mind describing those accounting tricks? I have a professional interest in them.
R&D wouldn't be my first choice if I were to make an *cof* accounting trick *cof*, because the rules bounding the line are too rigid and it tends to be highly scrutinized.

No specific answers to that, mostly because I know nothing about foreign accounting laws.
But just take a step back and look at current R&D spendings by the whole industry (doesn't apply to just this industry). Take away inflation, raises and a bit more and you still end up with massive increases in R&D spending that can be tough to explain with the amount of new 'highly skilled human resources' those companies seem to find. Also, some (mostly swiss) companies literally went bonkers in R&D spending a few years ago with barely any rampup time and no visible explanation. That triggered my alarm.

I don't think you understand how a new node is developed, which is probably why you can't believe the numbers.

My piece of the process development pie alone, just one little process development engineer, was roughly $8m/year for the fab work. Then there are the ex-fab costs, for example I had a $2m NSF federal grant that required dollar-for-dollar matching from TI over just 2yrs.

And I was just one engineer; I worked with literally hundreds of other engineers who all had to be equally resourced to get their job done. And even then, getting the job "done" was a four year journey for each node, longer still if you got into the path-finding research (like the NSF grant).

Thanks for the input. In this case I'm just standing in a distance, looking at numbers and doing educated guesses. What I noticed (and tried to convey) is that some people seem to shift R&D spending of companies to whatever argument they are fighting over without even a basic plausibility check (no offence, mrmt).
I did work with the R&D department of a small-ish company (~20.000 employees) until about 3 years ago. The company was mostly focused on a single product, but R&D were still churning out lots of unrelated projects. That shouldn't be something foreign to TSMC (and most certainly not Intel, IBM or for that matter TI) and can be a formidable money sink.

You mention tool vendors; I stand with my mouth open in aghast if you think tool vendors develop production worthy processes with their tools.

I would word it this way: They develop production worthy tools for predefined processes.

You don't get anything for nothing, if you think you are getting something for nothing then you are getting the wool pulled over your eyes. That is true in every industry, not just semiconductors.

There is a reason so many companies choose to be fabless. And the trend is decidedly uni-directional.

I get the impression that you rate the market harder than it is. Basic economics would make you believe that all markets will shrink to a monopoly, when in reality many markets stabilize or fade from relevance at some point before a monopoly happens.

edit: holy cow, what a long post

 

[mrmt]

bobcat infact it was 6 months early. Trinity was fine also its delay was llano inventory based. Vishera was fine as well. Jaguar also seems to be fine, Steam roller looks like it has slipped, the question is how much.

Bobcat started around 2005, and only in 2011 they got something to show, just when the netbook market crattered. Yes, right on time.

its not more inefficient, that's why they are using it, thats why GPU manufactures that design far bigger yet repeated ASIC's have been using them for many generations.

Oh, that's why CPU architects tend to shun synthesys designs.

you haven't actually described anything correctly...... s*** in s*** out. AMD's problems even during core2 era had nothing to do with R&D or money but direction. They had no plan and it showed.

Really? So what was Bulldozer canned in 2009? Do you know that for Bulldozer to be canned in 2009 in the ES stage it has to be started at least in 2006, no?

It was the stubborn insistence in a failed concept that brought AMD to where they are now, not that they were lacking direction.

 

[mrmt]

Those sums only add up if TSMC isn't spending anything at all on unrelated R&D and only throw out a single process every 2.5 - 3 years.

You are starting from an absurd proposition, that TSMC is capitalizing other expenses to R&D without them actually helping process development node and disregarding the most probable one, that all the R&D budget is process related. And the only objective evidence here is TSMC financial statements, which says that all R&D expenses are process related.

If you have any evidence that your proposition is true, feel free to share, if not, we can stay with TSMC' statements.

Underlining my point, CAPEX isn't R&D and should not be thrown into the same pot. But it is a much more valuable number to compare.

No, it's not underlining your point. CAPEX and R&D are *not* the same thing and they aren't being thrown in the same pot by me and by nobody here.

What I've been saying is that the problem is two-prongued, you have an R&D problem and you have a CAPEX problem, and it doesn't matter if you can get one but not the other, you have to sort out the two (I'm looking at you, GLF).

No specific answers to that, mostly because I know nothing about foreign accounting laws.

Foreign? Where are you living? You can throw IFRS or US GAAP here, I just want an accounting trick.

But just take a step back and look at current R&D spendings by the whole industry (doesn't apply to just this industry). Take away inflation, raises and a bit more and you still end up with massive increases in R&D spending that can be tough to explain with the amount of new 'highly skilled human resources' those companies seem to find. Also, some (mostly swiss) companies literally went bonkers in R&D spending a few years ago with barely any rampup time and no visible explanation. That triggered my alarm.

But that's exactly what the Rock's law said it would happen, right? Transistor budget grew exponentially, the complexity of each unit is bigger, testing and validation is a lot more complex... From the foundry POV, the margin for deviation is getting smaller and smaller, they have to cram more transistors in a very tiny space, etc. All points for more complex projects, and consequently more expensive projects.

You can't just throw Swiss companies (which companies? In which sector?) and then infer that there is an accounting fraud occurring in the MPU sector (which has no Swiss company in the first tier).

I did work with the R&D department of a small-ish company (~20.000 employees) until about 3 years ago. The company was mostly focused on a single product, but R&D were still churning out lots of unrelated projects. That shouldn't be something foreign to TSMC (and most certainly not Intel, IBM or for that matter TI) and can be a formidable money sink.

I don't think you get what R&D expense is.

R&D expense is *every* expense related to knowledge acquisition, including, but not limited to, materials, personnel, fees, equipment.

So if you have a foundry and pay a biologist to study butterflies in the parking lot of your company, then it is R&D. Maybe a bad R&D investment, but is still R&D. But on this same foundry, the expenses you need to develop your process node will be included on the same expense.

So if the R&D department of your company is researching other products or other subjects, something other than the product it should develop, it may be a bad management decision but it is still R&D expense. Only if the R&D department start to, let's say, develop labels, logos, or do accounting services, its expenses would have to be reclassified as SG&A.

This is the reason that the R&D number alone doesn't mean anything. They always get some specific topic on the financial statements. Have a peek at TSMC's or Nvidia (AMD is too opaque) annual report to see the R&D topics.

 

[Ajay]

Yes, if they will ramp 20nm in the same way as 28nm than they'll really can give up the foundry business.

I just remembered the FD-SOI story, there will be a presentation about it soon, hold by a GF guy:

http://www.soiconsortium.org/workshops/hsinchu/52c9e7312d249eda36071933123639a4/index.php

Seems GF will really use FD-SOI. If that works well, it could be a way out of the dilemma. If FD-SOI has more or less the same performance characteristics as 20nm bulk, they could position that against TSMC's 20nm.
AMD could be happy too, because they could port Kabini to it and thus fulfill their contracts with GF.

In short: The lateness of GF could be evened out a bit by the FD-SOI advantages.

The only problem is the state of FD-SOI ... we probably wont hear any interesting new at the presentation mentioned above, but lets wait and see ...

First, there better be some news at that event on FD-SOI, or everyone will just laugh at GFL and the SOI consortium. GFL's already dubious reputation will be destroyed as would their dreams of becoming a 1st tier global fab.

Secondly, the only reason that I can see for pushing 28nm at this point is that their 20nm node is in serious trouble and they need a stop gap. That also means that GFL will need FD-SOI for it's 20nm node since it'll be competing with TSMC's 16 nm FinFET design. That, or GFL will need to drop 20nm and go straight for their half baked 14XM node (and get no return on their 20nm R&D)

Speculating a bit, such a move would benefit AMD since it would become a standard process at GFL (no extra charges for R&D). If in fact AMD would likely breath a sigh of relief since their next gen Radeon GPU (if it is being made at GFL) could still compete with NV. Same for Kabini and whatever they have planed in 2014 (heck, Kabini on FD-SOI would save AMD a die shrink an they would still get a boost in power/performance). Steamroller Opterons on 28nm FD-SOI would be better than anything that was planned for 28nm Bulk. This could be good enough for them to be retain their position low cost alternative to Intel, instead of becoming completely irrelevant.

This wouldn't save AMD by any means, but it could give them some breathing room and save them some of the cost of designing components for a power/performance enhancements and a node shrink. And they might just require enough wafers to avoid paying GFL for nothing (WSA) - an odd situation were staying @ 28nm (larger dies) would become a net cost saver.

Given that nobody in the 1st tier chose to go GFL 28nm FD-SOI over TSMC 20nm, quite the opposite, everybody is lining up for 20nm, I guess we have an empirical evidence of what those numbers would show.

Doesn't matter what process parameter's GFL is selling. Their unpredictability is killing them (or maybe I should have said their predicable incompetence). It is interesting that one of the the items that TSMC made sure was highlighted, in the OP link, is the rate at which their process nodes are maturing - definitely a stab at GFL and a grab for more customers. Basically, that article said to me that TSMC want's to be the Intel of fab only pure plays. With Samsung having lost Apple (and because of the reasons they lost Apple), and all other fabs under-performing, they will become the Intel of pure-play fabs, barring a major screw-up.

 

[ShintaiDK]

I find it amazing anyone still has faith in GloFo. Its the same rerun over and over again. They lack money, they lack R&D, they lack customers.

Yet miracles are expected. Funny how its tied with a certain other company.

For fabless companies, anything but TSMC is irrelevant if they want to succeed.

 

[Ajay]

I find it amazing anyone still has faith in GloFo. Its the same rerun over and over again. They lack money, they lack R&D, they lack customers.

Yet miracles are expected. Funny how its tied with a certain other company.

For fabless companies, anything but TSMC is irrelevant if they want to succeed.

I have very little faith in GFL. If FD-SOI works for them @ 28nm, it's more of a stay of execution for AMD. If GFL's 20nm node is so screwed that they need FD-SOI, that is just horrible news the company.

1st and 2nd tier customers will lose all faith in the possibility of an independent foundry alternative to TSMC (much more expensive) that is not Samsung (less expensive, but they might steal your IP). Sucks for fabless semicon businesses who need a lower cost structure.

 

[AtenRa]

So in order to research 20nm *and* build a single factory with significant capacity you need some 12 billion dollars, spread around 4-5 years.

You are overbudget,

GlobalFoundries 22nm = 1.3B for Process Development + 6.7B for the fab = 8B

 

[Phynaz]

You are overbudget,

GlobalFoundries 22nm = 1.3B for Process Development + 6.7B for the fab = 8B

Undated slide and numbers from GF and SA for a process that hasn't been finished yet.

Chance of accuracy approaches zero.

 

[Piroko]

You are starting from an absurd proposition, that TSMC is capitalizing other expenses to R&D without them actually helping process development node and disregarding the most probable one, that all the R&D budget is process related. And the only objective evidence here is TSMC financial statements, which says that all R&D expenses are process related.
If you have any evidence that your proposition is true, feel free to share, if not, we can stay with TSMC' statements.

I'll give you that, they do say that all R&D expenses are process related. So my argument with unrelated R&D is unbased. Perhaps their investment into solar and lighting companies creates in-house R&D projects for older or different processes though.

Foreign? Where are you living? You can throw IFRS or US GAAP here, I just want an accounting trick.
You can't just throw Swiss companies (which companies? In which sector?) and then infer that there is an accounting fraud occurring in the MPU sector (which has no Swiss company in the first tier).

Granted, what made me listen up were companies in the healthcare sector, unrelated to the semiconductor industry. If I were able to name accounting fraud I certainly wouldn't be ranting about it here. I didn't even want to imply clear fraud, just that I'm sceptical how any single company can ramp up their R&D expenses by more than a billion (or, more than 20%) within a year without any aquisitions or otherwise noteworthy public announcements, all while other companies in the same sector do fine and hire as well. There is only so much supply in unemployed people with a specific academic degree. And only so much demand (/-variation) for relevant materials and equipment, at least compared to the sums we're talking about.

So if you have a foundry and pay a biologist to study butterflies in the parking lot of your company, then it is R&D. Maybe a bad R&D investment, but is still R&D. But on this same foundry, the expenses you need to develop your process node will be included on the same expense.

So if the R&D department of your company is researching other products or other subjects, something other than the product it should develop, it may be a bad management decision but it is still R&D expense. Only if the R&D department start to, let's say, develop labels, logos, or do accounting services, its expenses would have to be reclassified as SG&A.

This is the reason that the R&D number alone doesn't mean anything. They always get some specific topic on the financial statements. Have a peek at TSMC's or Nvidia (AMD is too opaque) annual report to see the R&D topics.

I concur absolutely, that was my main point. The R&D number is an intransparent figure without further information and can be split into any number of known and unknown projects which may or may not ever see a press release. From my personal experience there usually are a lot of unknown projects or stuff like that. Saw it driving at an event, that was funny.

 

[AtenRa]

Undated slide

Common Platform conference 2013

 

[Exophase]

1st and 2nd tier customers will lose all faith in the possibility of an independent foundry alternative to TSMC (much more expensive) that is not Samsung (less expensive, but they might steal your IP). Sucks for fabless semicon businesses who need a lower cost structure.

Not aware of Samsung ever stealing anything they got through their fabs. They were sued for making cell phones and tablets that, from what I understand, looked a tiny bit more like Apple's than those from every other cell phone and tablet manufacturer.

I don't understand why people are acting like the lawsuit between Apple and Samsung is scaring away any possible future fab business for Samsung. Most potential customers are not vertically integrated in the markets Samsung is, at least not to the extent Apple is.

They would be competing in SoC offerings, but that should bother Samsung more than it bothers those companies who'd want to make competing SoCs on their process. And yet I don't think it does bother Samsung, the same way ARM isn't bothered by selling licenses to make CPUs that compete with their own.

Qualcomm and nVidia were both complaining about limited capacity on TSMC 28nm, don't know how much that changed but Apple's move can't be helping. I could see both of the companies being more comfortable with Samsung than Apple currently is, will be interesting to see if they any business goes that way.

 

[AtenRa]

Well, maybe you're the guy to ask this one question: Is the Malta fab running production @ 28nm yet.

GLOBALFOUNDRIES Extending Fab 8 to Meet Strong Customer Demand

During the construction of Fab 8, we extended the shell of the Module 1 building with the expectation that our business would continue to grow. Today we see increasingly strong demand from our customers, especially at the 28nm node, and we are excited to be moving forward with this next phase in the development of the Fab 8 campus,” said Eric Choh, vice president and general manager, Fab 8, GLOBALFOUNDRIES

Common Platform conference 2013.

 

[ShintaiDK]

GloFo, slides and reality dont mix. Its a fact proven over and over, again and again.

 

[mrmt]

Granted, what made me listen up were companies in the healthcare sector, unrelated to the semiconductor industry. If I were able to name accounting fraud I certainly wouldn't be ranting about it here. I didn't even want to imply clear fraud, just that I'm sceptical how any single company can ramp up their R&D expenses by more than a billion (or, more than 20%) within a year without any aquisitions or otherwise noteworthy public announcements, all while other companies in the same sector do fine and hire as well.

My brother-in-law is a manager in one of the big four, the same I worked some time ago.

Whenever we talk about accounting, we sometimes share some story about customers, not disclosing the name of it for obvious reasons. As I worked in M&A and in corporate banking, I tend to shun small things to focus on the big picture. It might tell me something whether an expense is R&D or SG&A, but in the end it didn't really matter, because I was trying to look for money to repay a loan or to make the wallet of my client fatter with cash inflows, and an expense was reduced cash flow.

But to my brother-in-law, it did matter a lot what an expense was, because his function is to provide investors with the clearest picture of the current moment. His data won't be used for a single transaction as my data, he cannot count a data point as something that can be restructured or sold, he has to show what things really are. And some times this means huge reclassifications of expenses, or even adjustments in previous results, and loooong discussions with his customers.

What you are seeing in those companies might be the work of auditing teams recommending changes, or simply accounting directives from IAS/FASB that changed the accounting method of a given expense, or some clever accountant realized that someone was doing something wrong and they had to change accounting methods...

You may be right and those companies might have something wrong with it, but there are plenty of possibilities to check before assuming some accounting shenanigan.

 

[mrmt]

Undated slide and numbers from GF and SA for a process that hasn't been finished yet.

Chance of accuracy approaches zero.

If GLF is spending that paltry amount on R&D, that explains *a lot* of the state of affairs of their 28nm process.

Given four years to develop a new process, that puts them with some 300 million per year in R&D, while TSMC is spending 1,4 billion per year. Given that TSMC has at least two processes in the R&D pipeline, that would put TSMC R&D in roughly twice the size per node than GLF.

But this slide should be VERY outdated. No 28nm on it, no FDSOI, and GLF canned their 22nm process long time ago.

 

[Ajay]

GLOBALFOUNDRIES Extending Fab 8 to Meet Strong Customer Demand
Common Platform conference 2013.

Thank you for your post, but it doesn't answer my question, which is, "is Malta in volume production of 28nm wafers?". That chart isn't very useful in answering that question, and it appears to be from a SemiAccurate.com article about the decrease in tapeout @ GFL.

 

[Ajay]

Qualcomm and nVidia were both complaining about limited capacity on TSMC 28nm, don't know how much that changed but Apple's move can't be helping. I could see both of the companies being more comfortable with Samsung than Apple currently is, will be interesting to see if they any business goes that way.

Maybe 'stealing" is too harsh a word. But IDK has written some very good commentary about the way Asian companies look at competition compared to American and European companies. You may be able to find that info by searching.

TSMC: Well, if the rumors are true, Apple bought out the entire first quarter of non-risk 20nm wafer starts. Clearly, they want to get a jump on Qualcomm and they are using their substantial cash reserves to accomplish that. I expect that their 20nm ARM SoC design is likely to be their best yet (since they have probably spared no expense there as well). Apple's back is up against the wall and they are not comfortable there.

I was actually comfortable with my recent purchase of a GTX 680, knowing that Maxwell probably won't be available till 2H14 and may make for a nice Christmas present from wife when the time comes 😀

 

[itsmydamnation]

Bobcat started around 2005, and only in 2011 they got something to show, just when the netbook market crattered. Yes, right on time.

when was bobcats release ever on a road map ( your words) that it didn't meet. I have been following CPU/GPU arch for quite some time i never remember bobcat in 2005 care to provide a link. I remember only around 2008 was bobcat talked about and it wasn't on any road map at that point.

Oh, that's why CPU architects tend to shun synthesys designs.

no its not why CPU architects shun ( highly debatable to begin with) . Not my fault you choose the complete wrong metric to make a not very good point.

Really? So what was Bulldozer canned in 2009? Do you know that for Bulldozer to be canned in 2009 in the ES stage it has to be started at least in 2006, no?

yes and? How about go reread what i said, notice how you don't see the word bulldozer :whiste: .

It was the stubborn insistence in a failed concept that brought AMD to where they are now

How is CMT a failed concept, Intel failed at making a GPU several times that means GPU's are a failed concept, am I doing it right?

Please point out one thing where CMT itself is the issue. AMD's problem is their was very little evolution all the way from K8 to Llano in the CPU core itself. They got smashed by Core2 and failed to react, but it should have been apparent to AMD when Yonah was ahead in IPC that things could and would get rough, yet the didn't react they just continued on their merry path of not doing much. That was where there lack of direct showed, if they had direction at that time Barcelona would have looked a lot different.


edit: i mean 90nm Dothan not Yonnah

 

[Exophase]

Maybe 'stealing" is too harsh a word. But IDK has written some very good commentary about the way Asian companies look at competition compared to American and European companies. You may be able to find that info by searching.

I know a lot of Chinese firms have played loose with IP in the past (not to say they all do). But I have no reason to lump a South Korean firm into that mindset, much less a global behemoth like Samsung.

Plus why wouldn't that reasoning apply to TSMC too? Not sure who you'd consider closer to China in a comparison between Taiwan and S Korea.. but that's probably not a comparison that really needs to be made.

TSMC: Well, if the rumors are true, Apple bought out the entire first quarter of non-risk 20nm wafer starts. Clearly, they want to get a jump on Qualcomm and they are using their substantial cash reserves to accomplish that. I expect that their 20nm ARM SoC design is likely to be their best yet (since they have probably spared no expense there as well). Apple's back is up against the wall and they are not comfortable there.

Yeah, Apple's shaping up into something kind of scary.

.. pardon the personal quip on this, but it really eats me up inside seeing the best ARM design locked to one vendor's product ecosystem (and one I have no interest in supporting). I tend to root for ARM stuff for a whole variety of reasons, personal and otherwise, but in this case I'd probably rather see Intel with the best parts over Apple :/

 

[joshhedge]

.. pardon the personal quip on this, but it really eats me up inside seeing the best ARM design locked to one vendor's product ecosystem (and one I have no interest in supporting). I tend to root for ARM stuff for a whole variety of reasons, personal and otherwise, but in this case I'd probably rather see Intel with the best parts over Apple :/

Apples SoC dominance can only increase competition in the mobile market. It's a win-win situation for the end users no matter which operating system one choses to use.

 

[mrmt]

when was bobcats release ever on a road map ( your words) that it didn't meet. I have been following CPU/GPU arch for quite some time i never remember bobcat in 2005 care to provide a link. I remember only around 2008 was bobcat talked about and it wasn't on any road map at that point.

http://forums.macrumors.com/showthread.php?t=898409&page=25

Please point out one thing where CMT itself is the issue.

CMT is an issue when you have to shrink the number of units per core to have an economically viable chip.

CMT is an issue when you add yet another layer of cache management, let alont thread synchronization, power management, etc.

CMT is an issue when all those units in a core must sit idle because the other core is going turbo for another thread.

 

[Exophase]

Apples SoC dominance can only increase competition in the mobile market. It's a win-win situation for the end users no matter which operating system one choses to use.

I don't know about that. The more money a company has the more means they have to increase the gap between them and the competition. Which leads to making more money than the competition. That's the way the system works.

 

[Idontcare]

You are overbudget,

GlobalFoundries 22nm = 1.3B for Process Development + 6.7B for the fab = 8B

GloFo didn't design 45nm, 32nm or 22nm. They pay an annual fee to IBM to have access to the process tech developed by IBM in Fishkill. The eco-system employees who get assigned to fishkill (glofo's assignees in fishkill) are basically there for process tech transfer back to the production fabs and for keeping tabs on IBM's progress.

It is actually a huge point of friction and frustration amongst fab-club members as they originally thought their employees would have more creative control and input on the node's development. But IBM keeps tight control on who calls the shots at every step in the development cycle.

At any rate the cost to develop 32nm (or 22nm) is not what GloFo paid to IBM, they are paying a fraction of the actual development costs as the other fab-club members are also subsidizing the development costs.

If you want a more accurate picture of what node development costs then you need to look at TSMC's annual R&D budget and realize those are annual contributions made towards programs that span 4yrs.

And Capex is a bear. TSMC spends more than 2x on annual capex than GloFo's annual revenue. Leading edge fabs are expensive. Look at Intel's capex.

 

[code65536]

I know a lot of Chinese firms have played loose with IP in the past (not to say they all do). But I have no reason to lump a South Korean firm into that mindset, much less a global behemoth like Samsung.

Plus why wouldn't that reasoning apply to TSMC too? Not sure who you'd consider closer to China in a comparison between Taiwan and S Korea.. but that's probably not a comparison that really needs to be made.

I'm going to take a guess here: Samsung has its fingers in so many pies that, chances are, it's also your competitor, whereas TSMC is just a fab.

 

[Piroko]

If you want a more accurate picture of what node development costs then you need to look at TSMC's annual R&D budget and realize those are annual contributions made towards programs that span 4yrs.

But since they release new nodes in less than 4 years, there have to be two nodes (a full node and a half one) in R&D at most times, right?