For the 10nm node and beyond, transistor research efforts are focused on high mobility designs with Ge and III-V channel, reducing VDD supply voltage as well as the subthreshold slope in transistors and optimizing multi-Vt designs. Eventually, lateral finFETs built from silicon nanowires may be required. As previously reported in the post “Status update on logic and memory roadmaps,” the 14nm node (which imec calls the “N” node”) is in development today, heading toward early production in 2013/2014. That will be followed by the N10 node in production at the end of 2015 and beginning of 2016. Then N7 and N5 will follow in 2017 and 2019.
A detailed look at the likely roadmap for logic devices built on bulk silicon wafers using finFET technology was provided at the recent International Technology Forum for the press at imec in Leuven, Belgium in October. An Steegen, senior vice president process technology provided the overview, highlighting research underway for the 10nm, 7nm and 5nm nodes.
Steegen said power was a concern in both high performance logic devices, which are thermally limited, and in mobile devices, which are battery limited. “What we’ve been trying to do at all our technology nodes is to try to step down that power curve, mainly be trying to lower the Vdd,” she said. The trick, she said, was to lower the power, but still retain performance, and the best way to do that is to make the subthreshold slope of the device steeper. She said the target was 16mv/decade, which is the limit of conventional transistors.
There’s a tradeoff, however, in that reduced Vdd often means increased variability, depending on the threshold voltage of the devices. “On the High Vt device, when you go low Vdd, you’re so close to the threshold voltage of your device that the spread becomes enormous. What you could do is switch to a low Vt device. That’s better for variability but your leakage is going to increase,” she said.
At imec, their research is focused on bulk silicon finFETs (others are exploring fully depleted SOI) with a replacement gate and high k. Work is still underway on ways to best integrate a replacement metal gate, and on multi-Vt devices. “That’s still work we are executing,” she said. She noted that imec has worked on high-k metal (HKMG) gates for more than 15 years, and is now looking at how to implement a replacement metal gate on a finFET device and enginner the Vts. “Uou also need to make sure your reliability comes together,” she said.
One way in which they have enabled multi-VT tuning (range up to 600mV) is with controlled Al diffusion in metal gate stack. “You want to make sure you can offer the designers a low Vt device, a standard Vt device and a high Vt device, which means that you need to be able to tweak the materials in your gate to cover that entire range of Vt scaling,” Steegen said.
She added that, for the 10nm node, they are engineering an entire silicon finFET platform. “That means we work on every single module going in, from scaling to the N10 dimensions, control of the fin height, making sure you get conformal doping in the fin, and source/drain engineering because you still want to get some form of stress from your source/drain.”
For the 7nm node, Steegen believes the channel materials will need to be replaced with higher mobility materials, germanium for PFETs and InGaAs for NFETS. To integrate these materials, a technique known as Aspect Ratio Trapping (ART) is used. This technique uses high aspect ratio sub-micron trenches to trap threading dislocations, greatly reducing the dislocation density of lattice mismatched materials grown on silicon. “You have engineer your dislocations and defects,” Steegen said.
ART is shown to be very effective for a wide variety of materials including Ge, GaAs and InP. It has been combined with epitaxial lateral overgrowth to create long, 18 micron wide strips of low dislocation density material. ART has been used to integrate many types of Ge and III-V devices on silicon including GaAs MOSFETs, GaAs lasers, GaAs tunnel diodes and a silicon infrared imager chip with monolithically integrated Ge photodiodes.
For PFETs, the technique involved, “recessing the silicon, growing silicon germanium buffer back and then strained germanium on the top. The STI is then going to be recessed and you have a strained germanium fin on the top,” Steegen explained. The same integration scheme is used for NFETs, but “it’s a little more complex to try to get to a strained InGaAs NFET channel because the lattice mismatch with silicon is larger. You have to use more buffers here and go a little bit deeper to grow all these buffers through the trench,” she said. “Aspect ratio trapping makes sure all the defects — and you’re going to have them in that strained/relaxed buffer — are trapped at the sidewalls of the STI so that they don’t reach the top silicon.”
This added complexity appears to be worth the effort based on modeling, which shows a net gain of 25% more performance at constant power, compared to a bulk silicon finFETs. “There is still a lot of benefit you’re going to get in one node by replacing these channel materials,” Steegen noted.
Beyond 7nm, imec is looking at higher mobility materials such as graphene, and also looking at new device architectures such as tunnel FETS. “At this point, we are looking at germanium source tunnel FET to overcome the tunneling barrier with a lower bandgap material at the source,” Steegen said. “We truly want to try to break that 60mv/decade subthreshold slope.” She said lots of progress has been made but there was more work to do to understand band-to-band tunneling mechanisms. The team is also looking at “2D” materials such molybdenum disulfide and tungsten diselenide.
I think you mean 60mV/decade (at “room” temperature)
Pete,
Nice post, thanks.
I think you mean 60mV/decade rather than 16mV/decade as the limit subthreshold slope of conventional devices.