Nvidia’s CTO on the Way forward for Excessive-Efficiency Computing

ff10
So the place will we ff10 flip for future scaling? We ff10 are going to proceed to ff10 look to the third dimension. ff10 We’ve created experimental gadgets that ff10 stack atop one another, delivering ff10 logic that’s 30 to 50 ff10 p.c smaller. Crucially, the highest ff10 and backside gadgets are of ff10 the 2 complementary sorts, NMOS ff10 and PMOS, which might be ff10 the inspiration of all of ff10 the logic circuits of the ff10 final a number of a ff10 long time. We consider this ff10 3D-stacked complementary metal-oxide semiconductor (CMOS), ff10 or CFET (complementary field-effect transistor), ff10 would be the key to ff10 extending Moore’s Legislation into the ff10 following decade.

ff10 The Evolution of the Transistor ff10

ff10 Steady innovation is a vital ff10 underpinning of Moore’s Legislation ff10 , however every enchancment comes ff10 with trade-offs. To grasp these ff10 trade-offs and the way they’re ff10 main us inevitably towards 3D-stacked ff10 CMOS, you want a little ff10 bit of background on transistor ff10 operation.

ff10
Each metal-oxide-semiconductor field-effect transistor, or ff10 MOSFET, has the identical set ff10 of fundamental elements: the gate ff10 stack, the channel area, the ff10 supply, and the drain. The ff10 supply and drain are chemically ff10 doped to make them each ff10 both wealthy in cellular electrons ff10 (
ff10 n ff10 -type) or poor in them ff10 ( ff10 p ff10 -type). The channel area has ff10 the alternative doping to the ff10 supply and drain.

ff10
Within the planar model in ff10 use in superior microprocessors as ff10 much as 2011, the MOSFET’s ff10 gate stack is located simply ff10 above the channel area and ff10 is designed to venture an ff10 electrical area into the channel ff10 area. Making use of a ff10 big sufficient voltage to the ff10 gate (relative to the supply) ff10 creates a layer of cellular ff10 cost carriers within the channel ff10 area that permits present to ff10 stream between the supply and ff10 drain.

ff10
As we scaled down the ff10 traditional planar transistors, what machine ff10 physicists name short-channel results took ff10 heart stage. Mainly, the space ff10 between the supply and drain ff10 turned so small that present ff10 would leak throughout the channel ff10 when it wasn’t purported to, ff10 as a result of the ff10 gate electrode struggled to deplete ff10 the channel of cost carriers. ff10 To deal with this, the ff10 trade moved to a completely ff10 totally different transistor structure referred ff10 to as a
ff10 FinFET ff10 . It wrapped the gate ff10 across the channel on three ff10 sides to offer higher electrostatic ff10 management.

ff10 Intel launched its FinFETs ff10 in 2011, on the ff10 22-nanometer node, with the third-generation ff10 Core processor, and the machine ff10 structure has been the workhorse ff10 of Moore’s Legislation ever since. ff10 With FinFETs, we may function ff10 at a decrease voltage and ff10 nonetheless have much less leakage, ff10 lowering energy consumption by some ff10 50 p.c on the similar ff10 efficiency stage because the previous-generation ff10 planar structure. FinFETs additionally switched ff10 quicker, boosting efficiency by 37 ff10 p.c. And since conduction happens ff10 on each vertical sides of ff10 the “fin,” the machine can ff10 drive extra present by way ff10 of a given space of ff10 silicon than can a planar ff10 machine, which solely conducts alongside ff10 one floor.

ff10
Nonetheless, we did lose one ff10 thing in shifting to FinFETs. ff10 In planar gadgets, the width ff10 of a transistor was outlined ff10 by lithography, and subsequently it’s ff10 a extremely versatile parameter. However ff10 in FinFETs, the transistor width ff10 comes within the type of ff10 discrete increments—including one fin at ff10 a time–a attribute also known ff10 as fin quantization. As versatile ff10 because the FinFET could also ff10 be, fin quantization stays a ff10 major design constraint. The design ff10 guidelines round it and the ff10 need so as to add ff10 extra fins to spice up ff10 efficiency improve the general space ff10 of logic cells and complicate ff10 the stack of interconnects that ff10 flip particular person transistors into ff10 full logic circuits. It additionally ff10 will increase the transistor’s capacitance, ff10 thereby sapping a few of ff10 its switching pace. So, whereas ff10 the FinFET has served us ff10 effectively because the trade’s workhorse, ff10 a brand new, extra refined ff10 method is required. And it’s ff10 that method that led us ff10 to the 3D transistors we’re ff10 introducing quickly.

ff10 Within the RibbonFET, the gate ff10 wraps across the transistor channel ff10 area to reinforce management of ff10 cost carriers. The brand new ff10 construction additionally allows higher efficiency ff10 and extra refined optimization. ff10 Emily Cooper

ff10
This advance, the RibbonFET, is ff10 our first new transistor structure ff10 because the FinFET’s debut 11 ff10 years in the past. In ff10 it, the gate absolutely surrounds ff10 the channel, offering even tighter ff10 management of cost carriers inside ff10 channels that are actually shaped ff10 by nanometer-scale ribbons of silicon. ff10 With these nanoribbons (additionally referred ff10 to as
ff10 nanosheets) ff10 , we are able to ff10 once more range the width ff10 of a transistor as wanted ff10 utilizing lithography.

ff10
With the quantization constraint eliminated, ff10 we are able to produce ff10 the appropriately sized width for ff10 the appliance. That lets us ff10 steadiness energy, efficiency, and price. ff10 What’s extra, with the ribbons ff10 stacked and working in parallel, ff10 the machine can drive extra ff10 present, boosting efficiency with out ff10 growing the world of the ff10 machine.

ff10
We see RibbonFETs as the ff10 most suitable choice for increased ff10 efficiency at affordable energy, and ff10 we will probably be introducing ff10 them in 2024 together with ff10 different improvements, reminiscent of PowerVia, ff10 our model of
ff10 bottom energy supply ff10 , with the Intel 20A ff10 fabrication course of.

ff10 Stacked CMOS

ff10 One commonality of planar, ff10 FinFET, and RibbonFET transistors ff10 is that all of ff10 them use CMOS expertise, which, ff10 as talked about, consists of ff10 ff10 n ff10 -type (NMOS) and ff10 p ff10 -type (PMOS) transistors. CMOS logic ff10 turned mainstream within the Nineteen ff10 Eighties as a result of ff10 it attracts considerably much less ff10 present than do the choice ff10 applied sciences, notably NMOS-only circuits. ff10 Much less present additionally led ff10 to better working frequencies and ff10 better transistor densities.

ff10
Up to now, all CMOS ff10 applied sciences place the usual ff10 NMOS and PMOS transistor pair ff10 facet by facet. However in ff10 a
ff10 keynote ff10 on the ff10 IEEE Worldwide Electron Units Assembly ff10 (IEDM) in 2019 ff10 , we launched the idea ff10 of a 3D-stacked transistor that ff10 locations the NMOS transistor on ff10 high of the PMOS transistor. ff10 The next yr, at IEDM ff10 2020, we introduced ff10 the design for the primary ff10 logic circuit utilizing this 3D ff10 approach ff10 , an inverter. Mixed with ff10 acceptable interconnects, the 3D-stacked CMOS ff10 method successfully cuts the inverter ff10 footprint in half, doubling the ff10 world density and additional pushing ff10 the bounds of Moore’s Legislation.

ff10 3D-stacked CMOS places a PMOS ff10 machine on high of an ff10 NMOS machine in the identical ff10 footprint a single RibbonFET would ff10 occupy. The NMOS and PMOS ff10 gates use totally different metals. ff10 Emily Cooper

ff10
Benefiting from the potential advantages ff10 of 3D stacking means fixing ff10 numerous course of integration challenges, ff10 a few of which is ff10 able to stretch the bounds ff10 of CMOS fabrication.

ff10
We constructed the 3D-stacked CMOS ff10 inverter utilizing what is named ff10 a self-aligned course of, by ff10 which each transistors are constructed ff10 in a single manufacturing step. ff10 This implies developing each
ff10 n ff10 -type and ff10 p ff10 -type sources and drains by ff10 epitaxy—crystal deposition—and including totally different ff10 steel gates for the 2 ff10 transistors. By combining the source-drain ff10 and dual-metal-gate processes, we’re capable ff10 of create totally different conductive ff10 varieties of silicon nanoribbons ( ff10 p ff10 -type and ff10 n ff10 -type) to make up the ff10 stacked CMOS transistor pairs. It ff10 additionally permits us to regulate ff10 the machine’s threshold voltage—the voltage ff10 at which a transistor begins ff10 to change—individually for the highest ff10 and backside nanoribbons.

ff10
How can we do all ff10 that? The self-aligned 3D CMOS ff10 fabrication begins with a silicon ff10 wafer. On this wafer, we ff10 deposit repeating layers of silicon ff10 and silicon germanium, a construction ff10 referred to as a superlattice. ff10 We then use lithographic patterning ff10 to chop away elements of ff10 the superlattice and depart a ff10 finlike construction. The superlattice crystal ff10 gives a robust help construction ff10 for what comes later.

ff10
Subsequent, we deposit a block ff10 of “dummy” polycrystalline silicon atop ff10 the a part of the ff10 superlattice the place the machine ff10 gates will go, defending them ff10 from the following step within ff10 the process. That step, referred ff10 to as the vertically stacked ff10 twin supply/drain course of, grows ff10 phosphorous-doped silicon on each ends ff10 of the highest nanoribbons (the ff10 long run NMOS machine) whereas ff10 additionally selectively rising boron-doped silicon ff10 germanium on the underside nanoribbons ff10 (the long run PMOS machine). ff10 After this, we deposit dielectric ff10 across the sources and drains ff10 to electrically isolate them from ff10 each other. The latter step ff10 requires that we then polish ff10 the wafer right down to ff10 good flatness.

ff10 An edge-on view of the ff10 3D stacked inverter exhibits how ff10 difficult its connections are. ff10 Emily Cooper

ff10 By stacking NMOS on high ff10 of PMOS transistors, 3D stacking ff10 successfully doubles CMOS transistor density ff10 per sq. millimeter, although the ff10 true density is determined by ff10 the complexity of the logic ff10 cell concerned. The inverter cells ff10 are proven from above indicating ff10 supply and drain interconnects [red], ff10 gate interconnects [blue], and vertical ff10 connections [green].

ff10
Lastly, we assemble the gate. ff10 First, we take away that ff10 dummy gate we’d put in ff10 place earlier, exposing the silicon ff10 nanoribbons. We subsequent etch away ff10 solely the silicon germanium, releasing ff10 a stack of parallel silicon ff10 nanoribbons, which would be the ff10 channel areas of the transistors. ff10 We then coat the nanoribbons ff10 on all sides with a ff10 vanishingly skinny layer of an ff10 insulator that has a excessive ff10 dielectric fixed. The nanoribbon channels ff10 are so small and positioned ff10 in such a method that ff10 we are able to’t successfully ff10 dope them chemically as we’d ff10 with a planar transistor. As ff10 an alternative, we use a ff10 property of the steel gates ff10 referred to as the work ff10 operate to impart the identical ff10 impact. We encompass the underside ff10 nanoribbons with one steel to ff10 make a
ff10 p ff10 -doped channel and the highest ff10 ones with one other to ff10 type an ff10 n ff10 -doped channel. Thus, the gate ff10 stacks are completed off and ff10 the 2 transistors are full.

ff10
The method might sound advanced, ff10 but it surely’s higher than ff10 the choice—a expertise referred to ff10 as sequential 3D-stacked CMOS. With ff10 that technique, the NMOS gadgets ff10 and the PMOS gadgets are ff10 constructed on separate wafers, the ff10 2 are bonded, and the ff10 PMOS layer is transferred to ff10 the NMOS wafer. Compared, the ff10 self-aligned 3D course of takes ff10 fewer manufacturing steps and retains ff10 a tighter rein on manufacturing ff10 value, one thing we demonstrated ff10 in analysis and reported at ff10 IEDM 2019.

ff10
Importantly, the self-aligned technique additionally ff10 circumvents the issue of misalignment ff10 that may happen when bonding ff10 two wafers. Nonetheless, sequential 3D ff10 stacking is being explored to ff10 facilitate integration of silicon with ff10 nonsilicon channel supplies, reminiscent of ff10 germanium and III-V semiconductor supplies. ff10 These approaches and supplies could ff10 turn out to be related ff10 as we glance to tightly ff10 combine optoelectronics and different features ff10 on a single chip.

ff10 Making all of the wanted ff10 connections to 3D-stacked CMOS is ff10 a problem. Energy connections will ff10 should be constituted of under ff10 the machine stack. On this ff10 design, the NMOS machine [top] ff10 and PMOS machine [bottom] have ff10 separate supply/drain contacts, however each ff10 gadgets have a gate in ff10 widespread. ff10 Emily Cooper

ff10
The brand new self-aligned CMOS ff10 course of, and the 3D-stacked ff10 CMOS it creates, work effectively ff10 and seem to have substantial ff10 room for additional miniaturization. At ff10 this early stage, that’s extremely ff10 encouraging. Units having a gate ff10 size of 75 nm demonstrated ff10 each the low leakage that ff10 comes with glorious machine scalability ff10 and a excessive on-state present. ff10 One other promising signal: We’ve ff10 made wafers the place the ff10 smallest distance between two units ff10 of stacked gadgets is simply ff10
ff10 55 nm ff10 . Whereas the machine efficiency ff10 outcomes we achieved will not ff10 be information in and of ff10 themselves, they do evaluate effectively ff10 with particular person nonstacked management ff10 gadgets constructed on the identical ff10 wafer with the identical processing.

ff10
In parallel with the method ff10 integration and experimental work, we ff10 have now many ongoing theoretical, ff10 simulation, and design research underway ff10 seeking to present perception into ff10 how finest to make use ff10 of 3D CMOS. By means ff10 of these, we’ve discovered a ff10 few of the key concerns ff10 within the design of our ff10 transistors. Notably, we now know ff10 that we have to optimize ff10 the vertical spacing between the ff10 NMOS and PMOS—if it’s too ff10 quick it’ll improve parasitic capacitance, ff10 and if it’s too lengthy ff10 it’ll improve the resistance of ff10 the interconnects between the 2 ff10 gadgets. Both excessive ends in ff10 slower circuits that eat extra ff10 energy.

ff10
Many design research, reminiscent of ff10 one by
ff10 TEL Analysis Middle America introduced ff10 at IEDM 2021 ff10 , give attention to offering ff10 all the required interconnects within ff10 the 3D CMOS’s restricted house ff10 and doing so with out ff10 considerably growing the world of ff10 the logic cells they make ff10 up. The TEL analysis confirmed ff10 that there are numerous alternatives ff10 for innovation find one of ff10 the best interconnect choices. That ff10 analysis additionally highlights that 3D-stacked ff10 CMOS might want to have ff10 interconnects each above and under ff10 the gadgets. This scheme, referred ff10 to as ff10 buried energy rails ff10 , takes the interconnects that ff10 present energy to logic cells ff10 however don’t carry knowledge and ff10 removes them to the silicon ff10 under the transistors. Intel’s PowerVIA ff10 expertise, which does simply that ff10 and is scheduled for introduction ff10 in 2024, will subsequently play ff10 a key function in making ff10 3D-stacked CMOS a industrial actuality.

ff10 The Way forward for Moore’s ff10 Legislation

ff10 With RibbonFETs and 3D CMOS, ff10 we have now a transparent ff10 path ff10 to increase Moore’s Legislation ff10 past 2024. In a ff10 2005 interview ff10 by which he was ff10 requested to replicate on what ff10 turned his legislation, Gordon Moore ff10 admitted to being “periodically amazed ff10 at how we’re capable of ff10 make progress. A number of ff10 instances alongside the best way, ff10 I believed we reached the ff10 tip of the road, issues ff10 taper off, and our inventive ff10 engineers provide you with methods ff10 round them.”

ff10
With the transfer to FinFETs, ff10 the following optimizations, and now ff10 the event of RibbonFETs and ff10 finally 3D-stacked CMOS, supported by ff10 the myriad packaging enhancements round ff10 them, we’d wish to suppose ff10 Mr. Moore will probably be ff10 amazed but once more.