20nm is about 24% more expensive per unit area and with 1.9x more die gives a 35% cost reduction. The problem is at 16nm where the 1.05x more die than 20nm will not offest the increased cost per unit area and the cost per die will go up.
Wait for carbon tubes to redesign this roadmap. This obsession with smaller feature size is not a roadmap for long haul, it is a marketing message to show innovation. I do not see that value - moving from 22nm to 16nm. I know it was a big deal from 3.3um to 90nm
Compared to 20HKMG, 16FinFET's value is 20% better speed or 35% less power consumption. There's no density advantage migrating from 20nm to 16nm and the Cost per area/die will go up at 16nm. All supply chain members must implement continuous cost cutting program to accommodate the higher cost.
It's a good idea for tsmc by achieving 16nm development with 20nm development as a stepping stone. Combining too many revolutionary BEOL (metal double patterning) and FEOL (FinFET) variables in one development may result in much longer development time.
Significantly revised design flows with so many new design rules to follow (65nm with 800 rules, 20k deck size; 20nm 3k, 80k; and 16nm 3.4k, 100k), so many new simulation corners to signoff (28HPM at 280 corners and 16FinFET 360 corners) and so many analog circuit to redesign (100% redesign at 16FinFET), it's good to have two nodes for designers to migrate smoothly.
It's a big waste if the complicated gate-last HKMG technology has been used in only one node at 28nm, especially for tsmc with overwhelming market share win at the HKMG node.
TSMC's gate last HKMG will continue to be used at 20nm and 16nm, it wasn't just for 28nm.
Additional comments relevent to this thread:
The key issue right now it the lateness of EUV which is making shrinks much harder and more expensive. TSMC's 20nm will be a full shrink with some cost reduction, 16nm is essentially no shrink and a big increase in die cost so only customers who need performance are likely to use it. At 10nm - hopefully with EUV we should see a full shrink again with a full cost reduction.
It does look like Multigate is the leading technology for 16nm/14nm and below. FDSOI/ETSOI appears to be a low power/RF niche technology to me.
Please note I say multigate and not finFET. A classic finFET is a fin with gates on both sides. Intel's Trigate is a fin with gates on the top and both sides, that is a significant difference. For Trigate you have to deal with the high electric field at the top corners but you also get to relax the fin width.
For fully depleted operation SOI silicon thickness has to be 1/3 the gate length and that is really hard to do as gate lengths scale down. For finFETs with 2 gates the fin width must be 1/2 the gate length. For TriGate the fin width is 1x the gate length making manufacturing easier.
It will be interesting to see whether TSMC makes a classic 2 gate finFET or a 3 gate Trigate.
If the quadruple patterning is double the cost but 4x the density from single patterning, it still continues cost reduction. Even with EUV, the energy cost is high, so there is no benefit at that point over multiple patterning.
The cost of EUV is much lower than the cost of quadruple patterning at any reasonable throughput.
The energy required for an EUV exposure is more than the energy for a ArFi exposure but is less than the two ArFi exposures plus depositions and etches required for double patterning and is much lower than what is required for quadruple patterning.
Quadruple patterning at 10nm could require up to five cut masks and that would be a huge cost problem.
I am not sure where you get the EUV energy (dose) estimates. What you see in publications or the reports from ASML or IMEC (15-20 mJ/cm2) is not going to work in the high-volume manufacturing, with billions of features requiring a certain dose precision. More like 60 and going higher. It makes no practical difference from multiple patterning because, for a smaller node, this dose has to go up, similar impact to increasing ArFi exposures.
The cost of EUV is much lower than the cost of quadruple patterning at any reasonable throughput.
At some point in the future the cost of EUV could be cheaper than quadruple patterning. But when will that be? What throughput is reasonble and what will it take to get there? I would not make such a blanket statement today. The whole EUV infrastructure will still be quite expensive in the near term. And I think there are still too many unknowns today to declare that EUV lithography will be cost competitive by a certain date. I don't think its guaranteed. Obviously each company will have to decide what it's costs will be with either scenerio, which will depend on many factors. Whether or not EUV will be ready for a 10 nm node remains to be seen.
>> "For fully depleted operation SOI silicon thickness has to be 1/3 the gate length and that is really hard to do as gate lengths scale down. For finFETs with 2 gates the fin width must be 1/2 the gate length. For TriGate the fin width is 1x the gate length making manufacturing easier."
Could you elaborate in detail the relationship among gate length, fin width, and trigate / dual-gate? By "gate length", do you mean transistor width? Isn't "gate length" also known as channel length? Transistor width is a function of fin pitch, fin height, fin width, and whether the fin is tri- or dual- gate. Isn't it?
In figures of IvyBridge Tri-gate 22nm process reported by Chipworks: fin_width(bottom)~=15nm, fin_height~=34nm, and gate length 23.7~24.6nm. Fin_width(top) seems to be about of 1/2 of bottom.
Thanks for clarifying the meaning of gate length. I still don't understand: "For finFETs with 2 gates the fin width must be 1/2 the gate length. For TriGate the fin width is 1x the gate length making manufacturing easier." Could you describe further?
The whole idea behind FDSOI and Multi-Gate is to create a fully depleted channel in order to minimize off-state leakage.
Simply put a gate only controls a finite thickness.
For FDSOI you have one gate on top of the channel with a buried oxide underneath the channel. The silicon device layer thickness (the thickness of the film under the gate) must be 1/3 the gate length or the gate can't control the channel and achieve a fully depleted device. This is the major scaling challenge for FDSOI, it is very hard to make SOI device layers that thin.
For a FinFet you have a thin fin with two gates, one on each side "squeezing" the channel and the fin width (thickness) can be ½ the gate length and still achieve a fully depleted channel.
For a TriGate you have three gates squeezing the cahnnel from both sides of the fin and the top and the fin width (thickness) can be the same as the gate length and still achieve a fully depleted device.
Basically the more gates you have controlling the channel the thicker it can be and still be fully depleted in the off state.
There are even solutions being explored for the long term with gates on all four sides for even better control.
Please note that in each case the thickness can be smaller but not larger than the specified amount and achieve fully depleted operation.
Wondering if the power saving numbers are backed by Si or are just model numbers. Back in 2011 we heard claims of 50% power reduction and what was shown in product was about 20% at high end and almost no change for low power parts.
Doesn't look like they got anything at 20nm working above 3ghz, no gen2+ pcie or hdmi. Got to wonder what sort of pll their making, also. Guess they got three more months to actually sell a 20nm this year. BTW, no talk of yield?
Not surprising to see TSMC pushing forward to smaller technologies. With the number of wafers they process, they have the volume to push the envelope. The great thing is the number of companies that will benefit from this advancement.
DSA is still pretty early in the development process. It is a very interesting alternative but still a long way off.
The key question really is what throughput is EUV going to hit and when. In the simulations I have run EUV at 50 wph is already less expensive than triple pattering let alone quadruple patterning. ASML is forecasting 88 wph next year so if they meet that EUV could potentially be a 10nm alternative in a year.
I don't expect EUV to need double patterning at 10nm, certainly not for logic.
It might be currently 30 targeting 80 WPH for 15 mJ/cm2, which is the current reference dose for a lot of resist characterization. But if the minimum dose needs to be over 60 mJ/cm2, for a chip containing billions of 20 nm contacts or line cuts, then the WPH target is not met. The minimum dose doubles with each successive node.
Zeiss reported at SPIE 2013 the current EUV multilayer for the optics is limited near 13-14 nm half-pitch, which could be 7 nm or 10 nm logic or in between. So if this part of the infrastructure is not changed in time, it's a limit on EUV single exposure.
Unfortunately the past projections for EUV have always been overly optimistic, so it's difficult for me to believe that all of the known (let alone the ones we don't know about) barriers to successful cost effective implementation of volume manufacturing with EUV lithography are going to be resolved in the next two years.
It seems to me given the current 13.5 nm wavelength, the modest NA allowed by the current reflective optics, optical flare and shot noise, resist edge roughness, and other image degrading factors, EUV with single patterning will be lucky to achieve one useful node. Then what? Knowing what we know today we might have been better off taking the hit a couple of years ago and putting more resources into optimizing BEUV (beyond EUV, 6.7 nm wavelength). It's still possible that after all this time and money 13.5 nm wavelength EUV could ultimately be seen as limited just as 157 nm and X-ray lithography technologies were before and simply abandoned.
In all these preliminary design releases, has TSMC said anything specific on the methods used to dope the bulk finFETs, including the source/drain contants and extensnions and, most interesting, the junction isolation doping in the fin base regions? Another key item will be the methods for workfunction tuning of the HKMG stacks. So, do we wait for Dec and the IEDM talks & discusisons or is anything known already?
I haven't seen anything specific but TSMC tends to follow Intel's technology path. I would expect the fin Vt to be set by implant and in situ doped raised eSiGe and eSiC Source/Drains.
For what it is worth I expect EUV system performance NA, Flare, et . To improve over time and there to be multiple single pattern nodes.
An older ASML roadmap I saw included going to a ~6.8nm wavelength around 2018 but I am told by Litho friend that there are lot of issues with that wavelength.
Also I am told by my Litho expert friends that the statement earlier in this thread about dose doubling with each node is not correct.
Dose doubling comes from feature halving area each successive node but the number of photons cannot be halved, because that would increase the shot noise. The photon number should at least stay the same, but since the feature population is increasing as well, even 6 standard deviations allowed for this noise within a population of a billion may not be enough.
Ideally, the EUV wavelength would be continued with higher NA but this means larger angles, and the multilayer must be modified from its current status. It could impact the reflectivity and therefore the throughput. So there was also a proposal to increase the mask demagnification from 4X to 6X at least. That's another big change.