By Dick James, Senior Analyst, Chipworks
In my conference preview blog last week, I mentioned that session 3 on the Monday afternoon would be a hot session, with three finFET papers, by TSMC, Intel, and IBM. I was right – even though they were given in the Grand Ballroom, it was full.
Paper 3.1 from TSMC disclosed what looks like their 16FF+ 16nm finFET technology, advanced from the 16FF reported last year – although they don’t actually call it that in the paper. A 15% speed boost and 30% power reduction is claimed, or 40% speed gain and 60% power saving compared to the 20nm process.
Gossip in the industry has it that 16FF was not advanced enough for TSMC’s customers, so they did some transistor engineering and cranked up the performance; 16FF is not even mentioned on the website these days, and 16FF+ is now in risk production, with endorsements by Avago, Freescale, LG Electronics, MediaTek, Nvidia, Renesas and Xilinx.
The 48nm fin pitch and 90nm contacted gate pitch announced last year were maintained, as is the 1x metal pitch of 64nm. This level uses “advanced patterning scheme” – presumably self-aligned double patterning (SADP), whereas the other 80/90 nm pitch metals are done with single patterning. The low-k dielectric stack has been optimized relative to the 16FF process to give almost 10% capacitance improvement, and they have also added a planar high-k MIM capacitor (>15 fF/um2) for on-chip noise reduction.
At the transistor level, we have a dual-gate oxide process, replacement metal gate (gate-last), dual epitaxial raised source/drains, and tungsten local interconnect – but NO PICTURES! Lots of plots, but no transistor images, as in last year’s 16FF paper, and we were out of luck in the live presentation as well.
So we still have no idea of what the TSMC finFETs will look like. I guess that’s good for me and Chipworks, since we’ll have to wait until they actually show up in the real world sometime next year.
Intel gave a late news paper (3.7) describing their 14nm finFET (note – finFET, not trigate) process at 4.05 pm. Being late news, there were only 15 minutes for Sanjay Natarajan to describe what looks like a technology that is distinctly changed from the 22nm process. AND there were images!
As announced back in August, fin pitch is reduced to 42nm, contacted gate pitch to 70nm, and 1x metal to 52nm, and we confirmed these in our blog on the Broadwell chip that we pulled out of a Panasonic laptop. In addition to the fins, the gates and the minimum metal levels use SADP, making for complex front-end lithography.
The fins have been modified from the 22nm process to have a more vertical profile, slimmed down to 8nm wide, and Intel also claims a “novel sub-fin doping technique” using “solid-source doping to enable better optimization of punch-through stopper dopants.” Sanjay’s presentation revealed that the solid-source doping uses a doped glass; now it’s down to us to work out when and where it’s used for punch-through inhibition. Idsat is claimed to improve by 15% for NMOS and 41% for PMOS over 22nm, and Idlin by 30% for NMOS and 38% for PMOS.
Changes have also been made to the back end – low-k dielectrics are used in the first eight levels, and significantly we see the first use of air-gaps in the M4 and M6 levels (80 and 160nm pitch). This is Intel’s SEM image from the paper:
And here’s a TEM image from our analysis:
can see from the spacing of the gaps and the profile of the barrier layer over the copper that a patterned approach has been taken, as described in the IITC 2010 paper , using a mask step after the formation of the metal seal layer.
Intel likes to point out their history – this is the second generation finFET, fourth generation HKMG, and sixth generation strained silicon; will their 10nm be the third, fifth, and seventh generations?
I’m now inclined to think so, since at an Applied Materials event in the evening, when asked about the delay in the 14nm launch, Mark Bohr was heard to say “We won’t have similar problems at 10nm”. Mark does not make such comments lightly, so to me that implies two things – the 10nm process is pretty well locked down already, and it’s unlikely that there are huge structural changes from the 14nm generation. Indeed, the aggressive shrink from 22nm to 14nm puts them well on the way to the predicted 10-nm feature sizes.
Immediately after Intel’s talk IBM had their 15 minutes of IEDM advanced CMOS fame, describing their 14nm technology. This has their fourth generation embedded DRAM, but is the first-gen finFET, and the first-gen gate-last process (and I’ve lost count of the SOI generations).
IBM claims a “unique dual workfunction process applied to both NFETs and PFETs” and sub-20nm gate lengths, which will be the smallest we’ve seen if we ever get a sample. Being IBM, the intended product will be over 600 mm2 and have 15 metal levels, presumably their Power9 server chip.
Fin pitch is the same as Intel at 42nm, but contacted gate pitch is 80nm, and 1x metal is 64nm. Here the fins are completely isolated since they are on the buried oxide, so no punch-through implants are needed at the base of the fin as on a bulk silicon substrate.
We do have pictures – these are really fuzzy, but we can see the gate wrapped over the fin with slightly raised source/drains on either side, and some nice facets on the source/drain epi.
During the presentation there were (of course) no details of the work-function materials, but it was stated that two masks were used to make the dual work-function structure; so presumably two slightly different material sets for the different work-functions. Another tidbit was that the pass-gate transistors
In the e-DRAM had a different Vt than the logic transistors, but not achieved by a workfunction change.
I’d missed it, but the IBM alliance gave a paper at the VLSI conference back in June , where they describe a 10nm finFET process; this look likes the same process, backed off to 14nm and with the e-DRAM added.
The e-DRAM introduces some challenges in connecting the trench capacitor plate to the fin of the pass gate. In the planar 22nm version there is a polySi strap from the polySi in the trench to the SOI on the buried oxide; in the finFET design the polySi strap is still used, but it is formed as a plug on the trench fill connecting to the SOI layer before fin definition, and the plug is etched into a fin during the fin etch. The epi module has been tuned to minimise the strap resistance and therefore the effect on access time.
Cell size of the eDRAM is now 0.0174 μm2; and if the trench capacitors are coupled together without the select gates, they can provide on-chip decoupling capacitors with a value of 450 fF/um2.
In the back-end IBM has their fifteen layers of metal ranging from 1x – 40x, and the section shows that the 40x is seriously thick, to take the power needed to run a chip this size!
That made for an eventful afternoon, with a bit of a disappointment from TSMC; we’ll look forward to seeing both their finFET and the Power9 next year. Of course we have a suite of reports on the Intel Broadwell, for those who want a detailed analysis of the part!
 H.J. Yoo et al., “Demonstration of a reliable high-performance and yielding Air gap interconnect process”, IITC 2010, pp. 1-3
 K-I Seo et al., “A 10nm Platform Technology for Low Power and High Performance Application Featuring FINFET Devices with Multi Workfunction Gate Stack on Bulk and SOI”, VLSI Tech 14, pp. 12-13
Pingback: Laser spike annealing resolves sub-20nm logic device manufacturing challenges | Solid State Technology
IEDM – Monday was FinFET Day
Thanks to the authors of three papers at IEDM for their great contributions. May I receive electronic versions of the three paper on finFet. I wish also to receive their email addresses and phone numbers.
I wish to invite them to give presentation in IEEE seminars in Rochester, New York.
Thanks for your help.
Pingback: Blog review January 26, 2015 | Semiconductor Manufacturing & Design Community
Pingback: KLA-Tencor extends its 5D patterning control solution with new metrology systems | Solid State Technology
Pingback: KLA-Tencor extends its 5D patterning control solution with new metrology systems | Anchor Science LLC
“Intel likes to point out their history – this is the second generation finFET, fourth generation HKMG, and sixth generation strained silicon; will their 10nm be the third, fifth, and seventh generations?
I’m now inclined to think so, since at an Applied Materials event in the evening, when asked about the delay in the 14nm launch, Mark Bohr was heard to say “We won’t have similar problems at 10nm”. Mark does not make such comments lightly, so to me that implies two things – the 10nm process is pretty well locked down already, and it’s unlikely that there are huge structural changes from the 14nm generation. Indeed, the aggressive shrink from 22nm to 14nm puts them well on the way to the predicted 10-nm feature sizes.”
You still think these thoughts? ?
OK, that’s a cheap shot, but there’s a real issue behind them. Something has clearly gone terribly wrong, and it seems reasonable to ask what that is. the possibilities that suggest themselves are
– Intel was flat out lying. They started down that path when they said 14nm was going to be as smooth as every prior process, and they just couldn’t get out of the hole of promises once they’d fallen in. This satisfies our desire for villains, but I find it hard to take seriously.
– Intel tried to bundle too much new technology too fast. They want to (or feel they NEED TO, given the multi-front competition they are facing) roll out 10nm as GAA transistors or quantum well FETs, or using a III-V material, and the combination of all the changes they want at once has proved too much. In a sense the obvious thing to have done two or three years ago was to back off from multiple changes at once and try a more gradual process, but I assume every three months it looked like “just three months more and we can get it working” followed by a new showstopper as soon as that problem was fixed.
– Something unrelated to the tech per se, like inability, once it was working in the lab, to get partners like ASML or Applied Materials to provide some sort of specialized equipment in large enough quantities.
My guess is on the second option.
Intel, like MS, seems to like to operate on a model of bundling up many many changes to be released as an occasional big bang product. This might seem more efficient in some sense, and less disruptive — but it has the potential to go horribly wrong (we saw this with Longhorn and Vista).
TSMC, like Apple, seems to appreciate the dangers as tech becomes ever more complex, and so both seem willing to make annual incremental improvements that are never quite a singular big bang, but which also don’t risk a big blowout…
My guess (we shall see) is that post-10nm, we’ll see a rather more humble Intel switching, like MS has, to a schedule of more frequent, less dramatic process updates. More expensive and more disruptive in one sense, yes, but not as expensive as missing your very publicly announced and scheduled expected ship date by two years (or more…).
Wow. I had no idea chip manufacturing took place at this level. I guess it makes sense given how things continue to get both smaller and faster. Still pretty impressive though.
This is the best intuitive explanation for the FinFET I’ve seen so far.
Kudos to Threshold Systems.