PORTLAND, Ore.—Silicon wafer maker Soitec S.A. claims that chip makers can sidestep years of development work needed to perfect fully-depleted (FD) silicon transistors by switching to its silicon-on-insulator (SOI) wafers, a ploy that has already convinced STMicroelectronics NV, ST-Erikson and IBM Corp. to give it a try.
"Fully depleted transistor channels are quickly becoming a necessity for semiconductor manufacturers moving below the 32-nanometer node," said Steve Longoria, senior vice president of global strategic business development at Soitec. "IBM is going to SOI wafers for its FinFETs at the 14-nanometer node, and STMicroelectronics and ST-Erikson are working with us to develop 2D fully-depleted transistors for their next-generation mobile processors at 28-nanometer."
One of the biggest problems facing continued scaling of semiconductors below the 32-nanometer node is non-uniformities of dopants in the nanoscale-thin transistor-channel layer. To solve that problem, the industry is going to undoped channels for FD transistors. Intel has gone to great lengths to design FD undoped channels for its tri-gate FinFET transistor using standard bulk silicon wafers, which as a result requires sidewall implant doping to isolate the channel and prevent excess leakage current into the substrate.
Soitec has two flavors of its SOI wafers, one for traditional planar transistors that offers and ultra-thin top silicon layer with tollerances of plus or minus just five angstroms, for the FD transistor channels, atop an ultra-thin buried oxide layer that prevents leakage into the substrate without the extra process steps that Intel uses for its bulk-silicon process.
The second SOI wafer from Soitec is for 3-D FinFET transistors, such as those IBM has announced it will use at the 14-nm node. The 3-D SOI wafer has thicker top silicon layers for the tall 3-D fins and a thicker buried oxide layer to accommodate the higher fields produced by the multiple metal gates.
Both the 2-D planar and 3-D SOI wafers cost about four-times more than bulk silicon—accounting for Intel's reluctance to use them for its tri-gate FinFET process. But Soitec claims that the time gained in FD transistor development plus the fewer processing steps required when fabricating FD channels with sidewall implantation, more than makes up for the high price of the wafers.
"Our wafers cost about $500 wafer compared to $120 for bulk silicon," said Longoria. "But their price is recouped from process simplifications to give a three-to-10 times overall cost reduction."
Soitec claims FD transistors built using its SOI wafers provide 40 percent better performance, or because of the drastic cut in leakage current supplied by the buried oxide layer, can alternatively provide 40 percent lower power when operated at current performance levels. Soitec also claims to be working with both IBM and ARM to create a specification for SOI wafers for taking their traditional planar transistors to undoped FD channels, which can nix problems with short-channel effects where closely spaced source and drain electrodes start leaking through a bulk-silicon substrate.
As you know, Kris, major architectural choices are made years ahead of when we hear about them, and in the past the wafers had to be ready years ahead of the final choices. So that might have been a factor for Intel? And of course there's the question of where any given company put its priorities: margin, power, performance, etc.
But I was a little surprised that the news I'm hearing about Intel's getting 20% less power and 20% more performance with this new architecture. Somehow I thought they'd get more. What did you think?
IBM will do their FinFETs on SOI at 14nm, I understand. And they call their flavor of FD-SOI ET-SOI.
To the best of my knowledge, AMD has not actually given any indication of what they'll do at 20/22nm. (And remember that 28nm was *never* on the SOI-based part of their roadmap.) So maybe the jury's still out on that one.
But for sure there lots and lots of grumbling out there about yield & scaling right now. The FD (SOI) solutions (for both planar & 3D) take care of the variability problems that kill yield. Do you think that might sway a few votes?
Mark said that they made the choice for trigate back in 2008, when it became clear that the performance benefit from the fully depleted triple-gate structure (compared to 22-nm planar) was significant enough to justify the additional effort and cost of another step-function change in process architecture.
Compared with the 32-nm equivalent, the trigate gives a 37% performance increase at a lower voltage or a 50% power reduction at constant performance. Somehow Intel does this with no extra mask levels and only 2-3% additional cost (although extra litho steps are used, because of the need for double patterning).
thank you Adele, interesting blog...but as ST-E insists on advantages of fully depleted SOI over bulk CMOS it begs a question why others (I think Intel and AMd are in this camp) think the bulk CMOS is a way to go at 22nm...is it because of Fin-FET device? Kris
Check out ST-Ericsson's Technology blog -- just posted today. They really spell out (with a great chart) the advantage they're getting with 28nm FDSOI -- see http://blog.stericsson.com/blog/2012/04/st-ericsson-general/fd-soi-a-process-booster-for-st-ericssons-next-generation-novathor-part-1/.
Their conclusion: "So, over a large Vdd range (from 0.5V up to 1.3V), FD-SOI comprehensively outperforms existing bulk CMOS processes dedicated to mobile applications. This extra performance gain can be used either to increase peak performance or to operate at a lower Vdd for the same performance, saving dynamic power."
Things just got even more confusion. Up to this point, all SOI (on the digital side) has been partially-depleted. Now the industry's moving to fully depleted. Fully depleted can be done on bulk or SOI.
But, the only way to currently do fully depleted and keep the current planar ("2D") transistor approach is to use the ultra-thin SOI wafers. That's what STM and STE are doing at 28nm; the port to FD-SOI (as it's commonly called when it's 2D) will take their existing bulk designs to 14nm without upheaval, and they get really impressive results re: power/perf. They say it's particularly well-suited for the SOCs they're building for their target markets (smartphones, etc.)
FinFETs and other multigate devices are also fully depleted, and by virtue of their verticality are referred to as 3D (not to be confused with the whole TSV 3D thing!). FinFETs were invented on SOI, then Samsung found a way to do them on bulk -- but at a price in terms of the power/perf trade-off (if you do them on bulk, you have to dope, and then you get RDF, etc).
Doing FinFETs on SOI gives a big boost in terms of the power/perf advantages FinFETs promise (plus no doping). Soitec says it's also found that by using the SOI wafers for FinFETs, you save massively in terms of process steps. And they're also saying that those who opt for FinFETs on SOI should be able to cut their learning curve by about a year.
Lots more good info at the SOI Consortium website. See the FD-SOI Workshop presentations -- lots of detail on both planar & FinFET on SOI.
Looks like the article was edited and the portion about Intel's "reluctance" (without giving explaination as to why) adopting SOI was eliminated.
It's good to have some one like Mark Bohr on the team
BTW that's a fairly wide range
"But their price is recouped from process simplifications to give a three-to-10 times overall cost reduction."
I know that Varian is working with Soitec on SOI and a good portion of premium probably applies to higher cost.
I suppose IBM did not learn from the gate first versus gate last debacle
David Patterson, known for his pioneering research that led to RAID, clusters and more, is part of a team at UC Berkeley that recently made its RISC-V processor architecture an open source hardware offering. We talk with Patterson and one of his colleagues behind the effort about the opportunities they see, what new kinds of designs they hope to enable and what it means for today’s commercial processor giants such as Intel, ARM and Imagination Technologies.