Three-D-Stacked CMOS Takes Moore’s Legislation to New Heights

Possibly probably the most far-reaching technological fulfillment during the last 50 years has been the secure march towards ever smaller transistors, becoming them extra tightly in combination, and lowering their energy intake. And but, ever because the two people began our careers at Intel greater than two decades in the past, we’ve been listening to the alarms that the descent into the infinitesimal used to be about to finish. But 12 months after 12 months, sensible new inventions proceed to propel the semiconductor trade additional.

Alongside this adventure, we engineers needed to alternate the transistor’s structure as we persevered to scale down house and tool intake whilst boosting efficiency. The “planar” transistor designs that took us during the ultimate part of the twentieth century gave approach to Three-D fin-shaped gadgets by means of the primary part of the 2010s. Now, those too have an finish date in sight, with a brand new gate-all-around (GAA) construction rolling into manufacturing quickly. However we need to glance even additional forward as a result of our talent to scale down even this new transistor structure, which we name RibbonFET, has its limits.

So the place do we flip for long run scaling? We can proceed to appear to the 1/3 size. We’ve created experimental gadgets that stack atop each and every different, handing over common sense this is 30 to 50 p.c smaller. Crucially, the highest and backside gadgets are of the 2 complementary sorts, NMOS and PMOS, which might be the basis of all of the common sense circuits of the ultimate a number of a long time. We imagine this Three-D-stacked complementary metal-oxide semiconductor (CMOS), or CFET (complementary field-effect transistor), would be the key to extending Moore’s Legislation into the following decade.

The Evolution of the Transistor

Steady innovation is an crucial underpinning of Moore’s Legislation, however each and every growth comes with trade-offs. To grasp those trade-offs and the way they’re main us inevitably towards Three-D-stacked CMOS, you wish to have just a little of background on transistor operation.

Each metal-oxide-semiconductor field-effect transistor, or MOSFET, has the similar set of fundamental portions: the gate stack, the channel area, the supply, and the drain. The supply and drain are chemically doped to cause them to each both wealthy in cell electrons (
n-type) or poor in them (p-type). The channel area has the other doping to the supply and drain.

Within the planar model in use in complicated microprocessors as much as 2011, the MOSFET’s gate stack is located simply above the channel area and is designed to mission an electrical discipline into the channel area. Making use of a big sufficient voltage to the gate (relative to the supply) creates a layer of cell fee carriers within the channel area that permits present to glide between the supply and drain.

As we scaled down the vintage planar transistors, what instrument physicists name short-channel results took middle degree. Mainly, the space between the supply and drain turned into so small that present would leak around the channel when it wasn’t meant to, for the reason that gate electrode struggled to expend the channel of fee carriers. To deal with this, the trade moved to a completely other transistor structure known as a
FinFET. It wrapped the gate across the channel on 3 aspects to offer higher electrostatic regulate.

Intel presented its FinFETs in 2011, on the 22-nanometer node, with the third-generation Core processor, and the instrument structure has been the workhorse of Moore’s Legislation ever since. With FinFETs, shall we perform at a decrease voltage and now have much less leakage, lowering energy intake by means of some 50 p.c on the similar efficiency degree because the previous-generation planar structure. FinFETs additionally switched quicker, boosting efficiency by means of 37 p.c. And since conduction happens on each vertical aspects of the “fin,” the instrument can power extra present via a given house of silicon than can a planar instrument, which solely conducts alongside one floor.

On the other hand, we did lose one thing in transferring to FinFETs. In planar gadgets, the width of a transistor used to be outlined by means of lithography, and due to this fact this is a extremely versatile parameter. However in FinFETs, the transistor width comes within the type of discrete increments—including one fin at a time–a feature ceaselessly known as fin quantization. As versatile because the FinFET could also be, fin quantization stays an important design constraint. The design regulations round it and the need so as to add extra fins to spice up efficiency build up the full house of common sense cells and complicate the stack of interconnects that flip particular person transistors into whole common sense circuits. It additionally will increase the transistor’s capacitance, thereby sapping a few of its switching pace. So, whilst the FinFET has served us smartly because the trade’s workhorse, a brand new, extra delicate way is wanted. And it’s that way that led us to the Three-D transistors we’re introducing quickly.

Within the RibbonFET, the gate wraps across the transistor channel area to beef up regulate of fee carriers. The brand new construction additionally permits higher efficiency and extra delicate optimization. Emily Cooper

This advance, the RibbonFET, is our first new transistor structure because the FinFET’s debut 11 years in the past. In it, the gate totally surrounds the channel, offering even tighter regulate of fee carriers inside channels that at the moment are shaped by means of nanometer-scale ribbons of silicon. With those nanoribbons (also referred to as
nanosheets), we will be able to once more range the width of a transistor as wanted the use of lithography.

With the quantization constraint got rid of, we will be able to produce the correctly sized width for the applying. That shall we us steadiness energy, efficiency, and price. What’s extra, with the ribbons stacked and running in parallel, the instrument can power extra present, boosting efficiency with out expanding the realm of the instrument.

We see RibbonFETs as the most suitable choice for upper efficiency at affordable energy, and we will be able to be introducing them in 2024 together with different inventions, similar to PowerVia, our model of
bottom energy supply, with the Intel 20A fabrication procedure.

Stacked CMOS

One commonality of planar, FinFET, and RibbonFET transistors is that all of them use CMOS era, which, as discussed, is composed of n-type (NMOS) and p-type (PMOS) transistors. CMOS common sense turned into mainstream within the Nineteen Eighties as it attracts considerably much less present than do the other applied sciences, significantly NMOS-only circuits. Much less present additionally led to bigger running frequencies and better transistor densities.

So far, all CMOS applied sciences position the usual NMOS and PMOS transistor pair facet by means of facet. However in a
keynote on the IEEE World Electron Gadgets Assembly (IEDM) in 2019, we presented the concept that of a Three-D-stacked transistor that puts the NMOS transistor on most sensible of the PMOS transistor. The next 12 months, at IEDM 2020, we offered the design for the primary common sense circuit the use of this Three-D method, an inverter. Mixed with suitable interconnects, the Three-D-stacked CMOS way successfully cuts the inverter footprint in part, doubling the realm density and extra pushing the bounds of Moore’s Legislation.

Three-D-stacked CMOS places a PMOS instrument on most sensible of an NMOS instrument in the similar footprint a unmarried RibbonFET would occupy. The NMOS and PMOS gates use other metals.Emily Cooper

Benefiting from the prospective advantages of Three-D stacking approach fixing a lot of procedure integration demanding situations, a few of which can stretch the bounds of CMOS fabrication.

We constructed the Three-D-stacked CMOS inverter the use of what’s referred to as a self-aligned procedure, by which each transistors are built in a single production step. This implies establishing each
n-type and p-type resources and drains by means of epitaxy—crystal deposition—and including other steel gates for the 2 transistors. Through combining the source-drain and dual-metal-gate processes, we’re in a position to create other conductive kinds of silicon nanoribbons (p-type and n-type) to make up the stacked CMOS transistor pairs. It additionally permits us to regulate the instrument’s threshold voltage—the voltage at which a transistor starts to modify—one at a time for the highest and backside nanoribbons.

How can we do all that? The self-aligned Three-D CMOS fabrication starts with a silicon wafer. In this wafer, we deposit repeating layers of silicon and silicon germanium, a construction known as a superlattice. We then use lithographic patterning to chop away portions of the superlattice and depart a finlike construction. The superlattice crystal supplies a powerful toughen construction for what comes later.

Subsequent, we deposit a block of “dummy” polycrystalline silicon atop the a part of the superlattice the place the instrument gates will cross, protective them from the next move within the process. That step, known as the vertically stacked twin supply/drain procedure, grows phosphorous-doped silicon on each ends of the highest nanoribbons (the longer term NMOS instrument) whilst additionally selectively rising boron-doped silicon germanium at the backside nanoribbons (the longer term PMOS instrument). After this, we deposit dielectric across the resources and drains to electrically isolate them from one some other. The latter step calls for that we then polish the wafer all the way down to best flatness.

An edge-on view of the Three-D stacked inverter displays how difficult its connections are. Emily Cooper

Through stacking NMOS on most sensible of PMOS transistors, Three-D stacking successfully doubles CMOS transistor density in step with sq. millimeter, despite the fact that the true density relies on the complexity of the common sense mobile concerned. The inverter cells are proven from above indicating supply and drain interconnects [red], gate interconnects [blue], and vertical connections [green].

In any case, we assemble the gate. First, we take away that dummy gate we’d installed position previous, exposing the silicon nanoribbons. We subsequent etch away solely the silicon germanium, liberating a stack of parallel silicon nanoribbons, which would be the channel areas of the transistors. We then coat the nanoribbons on either side with a vanishingly skinny layer of an insulator that has a top dielectric consistent. The nanoribbon channels are so small and located in this type of approach that we will be able to’t successfully dope them chemically as we’d with a planar transistor. As an alternative, we use a assets of the steel gates known as the paintings serve as to impart the similar impact. We encompass the ground nanoribbons with one steel to make a
p-doped channel and the highest ones with some other to shape an n-doped channel. Thus, the gate stacks are completed off and the 2 transistors are whole.

The method would possibly appear advanced, nevertheless it’s higher than the other—a era known as sequential Three-D-stacked CMOS. With that means, the NMOS gadgets and the PMOS gadgets are constructed on separate wafers, the 2 are bonded, and the PMOS layer is transferred to the NMOS wafer. When compared, the self-aligned Three-D procedure takes fewer production steps and helps to keep a tighter rein on production price, one thing we demonstrated in analysis and reported at IEDM 2019.

Importantly, the self-aligned means additionally circumvents the issue of misalignment that may happen when bonding two wafers. Nonetheless, sequential Three-D stacking is being explored to facilitate integration of silicon with nonsilicon channel fabrics, similar to germanium and III-V semiconductor fabrics. Those approaches and fabrics might develop into related as we glance to tightly combine optoelectronics and different purposes on a unmarried chip.

Making all of the wanted connections to Three-D-stacked CMOS is a problem. Energy connections will wish to be produced from under the instrument stack. On this design, the NMOS instrument [top] and PMOS instrument [bottom] have separate supply/drain contacts, however each gadgets have a gate in not unusual.Emily Cooper

The brand new self-aligned CMOS procedure, and the Three-D-stacked CMOS it creates, paintings smartly and seem to have considerable room for additional miniaturization. At this early degree, that’s extremely encouraging. Gadgets having a gate duration of 75 nm demonstrated each the low leakage that incorporates very good instrument scalability and a top on-state present. Some other promising signal: We’ve made wafers the place the smallest distance between two units of stacked gadgets is solely
55 nm. Whilst the instrument efficiency effects we completed aren’t information in and of themselves, they do examine smartly with particular person nonstacked regulate gadgets constructed at the similar wafer with the similar processing.

In parallel with the method integration and experimental paintings, we have now many ongoing theoretical, simulation, and design research underway taking a look to offer perception into how very best to make use of Three-D CMOS. Via those, we’ve discovered one of the crucial key issues within the design of our transistors. Particularly, we now know that we wish to optimize the vertical spacing between the NMOS and PMOS—if it’s too brief it’s going to build up parasitic capacitance, and if it’s too lengthy it’s going to build up the resistance of the interconnects between the 2 gadgets. Both excessive leads to slower circuits that eat extra energy.

Many design research, similar to one by means of
TEL Analysis Middle The usa offered at IEDM 2021, center of attention on offering all of the vital interconnects within the Three-D CMOS’s restricted house and doing so with out considerably expanding the realm of the common sense cells they make up. The TEL analysis confirmed that there are lots of alternatives for innovation find the most productive interconnect choices. That analysis additionally highlights that Three-D-stacked CMOS will wish to have interconnects each above and under the gadgets. This scheme, known as buried energy rails, takes the interconnects that offer energy to common sense cells however don’t raise information and gets rid of them to the silicon under the transistors. Intel’s PowerVIA era, which does simply that and is scheduled for creation in 2024, will due to this fact play a key position in making Three-D-stacked CMOS a business fact.

The Long run of Moore’s Legislation

With RibbonFETs and Three-D CMOS, we have now a transparent trail to increase Moore’s Legislation past 2024. In a 2005 interview by which he used to be requested to replicate on what turned into his regulation, Gordon Moore admitted to being “periodically amazed at how we’re in a position to make growth. A number of occasions alongside the best way, I believed we reached the tip of the road, issues taper off, and our ingenious engineers get a hold of techniques round them.”

With the transfer to FinFETs, the following optimizations, and now the improvement of RibbonFETs and in the end Three-D-stacked CMOS, supported by means of the myriad packaging improvements round them, we’d love to assume Mr. Moore might be amazed another time.