Lets look at both of these as separate issues because in many ways they are not at all related. The first issue at hand is Intel’s lead on the rest of the semiconductor fabrication world, be they foundries or in-house producers. Intel is on 22nm, Intel is producing FinFETs in volume, and everyone else is two years behind. Make that two years behind getting to 22/20nm, closer to four years behind on making FinFETs, all taken with a large dollop of optimism.
The public roadmaps of the competition, that would mainly be TSMC and the Common Platform trio, Samsung, Global Foundries, and IBM, are going to go to 20nm with planar structures. Their roadmaps all point to FinFETs at the 14nm node, ostensibly 2 years behind 20nm. Publicly, 20nm will be here in 2013 although most industry watchers think 2014 is a much more likely possibility.
This is a massive lead for Intel, they have both a full shrink and a very difficult technical challenge, fins, developed and in mass production. No one else is even close. That lead however brings an even more thorny problem, tools. The tools needed for 22/20nm are expensive, hard to produce, and in demand from absolutely no one other than Intel. Therein lies the problem.
Why? If you are the tools industry, why would you want to sink the time, engineering bandwidth, and money in to a product line that has one customer? Granted, that customer is huge, has gobs of money, and is willing to buy bleeding edge products, but they aren’t stupid. Tools that go in to bleeding edge fabs are shatteringly expensive, price tags in the tens of million of dollars are not uncommon for litho equipment, and a modern fab needs a large number of them.
Even with this price and quantity, the revenue from a single massive client pales in comparison to the multi-year tools development process that each new node requires. To make a profit from the new tools that enable the next fabrication node, vendors need to have a lot of clients. In the past, this was the case, there would be a leader that bought the initial units, and by the time production ramped, others would be there. Some lagged, some were closer to the first, but there was a range of customer demand. The system worked.
Now, Intel is between 2-4 years ahead of the industry depending on how you count. If a vendor makes tools for Intel’s schedule, there is a long long lag between the time Intel buys their units and the day that the next order comes in. That type of lag is not desirable, and if you are allocating R&D dollars, tool purchases, and manufacturing lines, why be 2+ years ahead of the majority of your customer base?
That is why Intel’s lead is a blessing and a curse, they can produce transistors that no one else can, but they require tools that no one else,yet, needs to do it. If a vendor is going to produce those tools for Intel, Intel is probably going to be asked to pay more than their fair share or wait.
At the 22nm node, this didn’t bite Intel hard if at all. The litho tools are all use 193nm light sources, and the wafers are the same 300mm size they have always been. The new devices needed to enable 22nm are not all that different from the ones needed for 32nm, so it all worked out more or less to plan.
Unfortunately, there are two major transitions needed for post 20/22nm nodes, EUV and 450mm. Technically speaking, neither is absolutely necessary to make 14nm transistors, but cost and throughput concerns mean these new technologies are needed to do so profitably. The more commoditized the chips produced, the higher the pressure for more efficient tools at the next node. So the debate rages on as to where the exact break points are, and what the downsides are to not having one or both, but in short order, there will be no choice.
At SemiCon last year, all of the panels and talks on 450mm wafers said the same thing, the industry is moving there together. Some predicted this would be at 14nm, others are saying 10nm, but no one questioned that the major players would all have to move at once. That panel included high ranking Intel process personnel as well as other industry giants. If the industry decides that 14nm is the transition point for 450mm, everyone will be making 14nm chips on 450mm wafers, 10nm if that ends up being the consensus. What is clear is that no one will be doing a 450mm wafer a node earlier that the rest.
Why? Because the tools vendors all said that once the crossover point is agreed upon, they will only make tools for that node that use 450mm wafers, the market isn’t big enough to sustain both 300mm and 450mm variants of the same device. If Intel moves 2-4 years before everyone else, they will have to convince the entire tool chain to move just for them. To do that, all of Intel’s competition has to agree to do so as well. Possible? Yes. Probable? Not without guarantees of profitability for everyone.
The same holds true for EUV. While most think this will be necessary at the 14nm node, you can make do with 193nm light sources and triple pattering in a pinch. It may bite foundries in the margins, but it can be done. If EUV comes to market at the 14 or 10nm node, it doesn’t matter much because Intel will still be 2-4 years ahead, and face the same outstretched hands from the tools chain. Being a leader used to mean you could do what no one else could. Now it means you have to pay for what they don’t need yet to get the ball rolling.
What is the price tag for this lead? For lithographic equipment, the largest of the three vendors, ASML, needed a tad over $1 billion to grease the wheels. Of that, $680 million was for 450mm R&D, $340 was aimed at EUV development. While litho tools are probably the most complex parts of the chain, there are dozens of other steps that each require a new compound, tool, or process for the whole process to work. Even if everyone else together only costs what ASML did, that is not exactly a minor outlay for Intel.
If Intel doesn’t pay up, or one vendor can’t keep up for some reason, the rest are essentially paperweights. Second sourcing everything doesn’t really help because with one customer, that just means Intel is funding two competing R&D teams, and that just means twice the 10-digit price tag. That said, all this financial outlay will mean Intel can do what no one else can, and move to nodes before anyone else can too. The double edged sword is that they have to pay for what no one else needs, likely all by themselves.
To make matters worse, if, and when the rest of the players catch up, they will have a fully debugged set of tools to buy off the shelf, paid for with Intel sweat equity and dollars. Both are seen as much preferable to the foundries doing it themselves, but the ever-increasing Intel lead is not viewed in such a fond light.
Since Intel started mass production of 22nm wafers in late 2011, if their two year cadence on new nodes doesn’t change, you can reasonably expect that 14nm wafers will start in late 2013. That would be 18 months away. Tools orders would probably have to be done by now, and equipment moved in to the fabs starting in 1H 2013. It is not a quick process.
Since there aren’t really 450mm wafer tools available in mass quantities yet, nor is EUV ready for prime time, that precludes either tech for initial 14nm production. That leaves 10nm, or possibly an updated 14nm process for EUV, and orders need to be placed for those in about two years, give or take, with initial equipment installation in three or so. If these devices don’t exist yet, it is kind of hard to do either one.
So Intel is laying out $1 billion to make sure ASML is there. Other fabs are said to be interested in the concept, but for the moment, what’s their hurry? If Intel essentially has to pay for everything or twiddle their thumbs, guess what they will do. And what the others won’t, they can always jump on board when it suits their timetables.
On current nodes, there are enough fabs and foundries running wafers to justify tools development for the ‘mass’ market, misnomer that it may be. No one expects any new players to jump in, and everyone expects one or two marginal players to merge or shut down. The number of buyers for semiconductor equipment is shrinking fast.
With this shrinking pool of paying customers comes a need for vendors servicing them to have products that reach as much of the market as possible. This rush to the middle means no going out on a limb, no taking chances, and keeping in mind that a single misstep can end your company. If you go out on a limb, you do so knowing that there is either no cost in it, or have guaranteed returns.
That is why Intel paid ASML for EUV and 450mm development, $1 billion goes a long way towards meeting the no cost goal. That same money is a good indicator that Intel is both willing and able to buy the devices when they are ready, but it is not a guarantee. For those vendors who are not recipients of Intel’s largess, it is a time for some serious soul-searching.S|A
Latest posts by Charlie Demerjian (see all)
- SemiAccurate has Skylake-SP die shots, sizes, and more - Jul 21, 2017
- Skylake-SP has a diverged core - Jul 19, 2017
- Intel’s Purley platform architecture is a step forward - Jul 13, 2017
- Intel launches Purley aka metal Xeons - Jul 11, 2017
- AMD’s Epyc has lots of connections - Jun 26, 2017