We have a guest contribution today from Ze'ev Wurman, MonolithIC 3D Inc.'s Chief Software Architect. Ze'ev discusses about where is the semiconductor industry heading to.

The semiconductor industry is in the doldrums. The PC market shrinks, Intel shares sink, Applied Materials cuts staff, and even Apple suddenly experiences its share price drop by $100 in a month. Are things really so bad?

But other news seems different. TSMC shares are close to their historical high; Global Foundries leapfrogs TSMC technology and nips at Intel’s heels with 14nm; Samsung reports record quarterly profit. Things seem to be going swimmingly.

Semiconductor business has had its ups and downs since its inception. As demand followed, more and more capacity was put on line, which caused the next overcapacity and slump, inevitably followed by the next spike in demand as the technology inexorably marched down the scaling curve. So, perhaps, nothing is really new here after all.

Yet, perhaps, we should not be so sanguine anymore. We did cope with 193nm light to define our chips down to 20 nm, but at an ever-increasing cost of expensive phase shift masks, immersion lithography, and double exposure. EUV has been talked about for at least 15 years (following another 15 years of x-ray lithography development fiasco) and has been “late” since at least 2005. Despite the impressive progress shown by ASML, the industry greats – Intel, Samsung, and TSMC --banding around EUV is possibly more a sign of desperation than a strong vote of confidence. The drastic reduction in foundry players – from 20 or so in 90nm to four or less for 14nm has been noticed by many and cannot be good for the long-term health and vitality of the industry. 

ASML is the last remaining game in town and, even if it works, building a foundry at over $10B a pop, and developing a technology node at well north of $1B, does wonders at keeping everyone but the most committed (and with deep pockets) out of the game. And the unanswered question is still in place: at what price point will the industry effectively become the domain of the few mass-produced designs such as Apple’s or Samsung’s phones and nothing else? 

Clearly pressures have been building up and the industry can’t pretend everything is as usual for much longer.  3D devices has been talked about for decades, yet implementing this dream was considered infeasible until recently. 3D designs are, in a sense, the holy grail of the industry:

• They allow for shrinkage of the average source to destination distances, shrinking power dissipation and improving performance;
  

• They can include inexpensive built-in fault detection and repair as described here, allowing for as large as needed yielded dies, which cuts further on power by saving on off-chip I/O power;
• They allow cheap and high performance integration of dies with disparate technologies, so advanced (and expensive)  logic can be stacked with reduced-cost technology for memory dies, or specialized analog and RF functions;
• They allow the reuse of older fab lines as much of density improvement is achieved through stacking dies rather than shrinking features;
  

• They allow efficient heat removal without exotic cooling technologies through use of power delivery networks, to be presented in the upcoming IEDM, paper 14.2 ;
  

In recent years TSVs started to show up. Yet, while TSVs are good for designs that need limited vertical connectivity between disparate sub-systems such as processors with memory, they do not really open the door to a true monolithic 3D design.  What true monolithic devices offer is a much higher vertical connectivity, by a factor of up to 10,000, and enable the stacking of multiple dies.

The impact of the increased monolithic vertical integration at lower cost can be dramatic on every electronic market segment.

For mobile devices, the inexpensive integration of analog, RF, and sensors, can lower their cost and power consumption for an even broader market penetration and longer battery life.

In medicine, the footprint of devices is often critical. The availability of camera-in-a-pill, or of implantable medical devices that control drug release, improve hearing, monitor vital signs, or allow artificial vision, are all strongly dependent on heterogeneous device integration in a small footprint and with reduced power. Monolithic 3D is key to transforming the planar and bulky designs of today’s 2D to grain-of-corn and grain-of-rice shape factors that can be inserted for very long time periods into our bodies.

Fostering innovation and reducing barriers to entry of new products are considered crucial for future economic prosperity. FPGAs have been trying to fill this niche since the demise of ASICs, yet they suffer from many handicaps: they are physically large, they are power hungry, and they are available in a limited number of configurations that are often suboptimal for the application. Monolithic 3D technology allows the inexpensive creation of a nearly infinite number of FPGA configurations that can be tailored to every application, as described here. And it does so while dramatically reducing both the device footprint and power.

Large-scale computing is facing enormous challenges to reduce its power consumption. Server farms of the likes of Google, Amazon, or Facebook consume tens and hundreds of megawatts of energy, while the government struggles mightily to keep its planned Exascale supercomputer under 20MW. Three-dimensional chips can play a large role in reducing power consumption by reducing the interconnect length (and hence, its capacitance), which is responsible for most of the power dissipation in modern chips. Ultra large scale integration with high yields, enabled through 3D repair structures, will further slash the power that today resides in the off-chip drivers.

In memory design, the transition to 3D technology is already taking place as described in our previous blog. Monolithic 3D structures using crystalline silicon may further the penetration and efficacy of this technology in both non-volatile memory as well as in DRAM.

The semiconductor world will inevitably move to monolithic three-dimensional technologies. The change drivers are already here: the skyrocketing cost of scaled-down lithography, the need to reduce power dissipation, and the need for heterogeneous integration. The only question is how quickly it will move there, and who will be the winners.

We have a guest contribution today from Israel Beinglass, the CTO of MonolithIC 3D Inc. Israel discusses the opportunities of 3D NAND with Monolithic 3D.

NAND technology, which is a subset of NVM (Non Volatile Memory), was invented by Fujio Masuoka of Toshiba back in 1984. Flash memory was presented at IEDM1984 by Dr. Masuoka and his colleagues [1].  The following is a short quote from the original paper “the cell is programmed by a channel hot carrier injection mechanism similar to EPROM. The contents of all memory cells are simultaneously erased by using field emission of electrons from a floating gate to an erased gate in a FLASH (Hence the name FLASH)”.  

Masuoka came back to the IEDM in 1987 and suggested a Flash NAND structure [2].
Intel created the first commercial NOR type of Flash chips in 1988. For the next few years some major developments occur in the Flash arena:

• In 1989, Samsung and Toshiba created a NAND flash memory.
• In 1994, Compact Flash was invented and introduced by SanDisk.
• In 1999, the SD memory card was released by a combination of SanDisk, Toshiba and Matsushita.
• In 2001, the world’s first 1 Gigabit Compact Flash card was introduced.

From 2006 onwards, NAND became the most scaled of devices beating out the microprocessor devices (see Figure 1). The current state of the art is 20nm (2x) technology, as the world’s appetite for storage is still strong. Flash Cards, SSD, Smartphone and Tablets are the leading growing applications.

NAND memory as a true cross point array with the control gate on top of the floating gate and only one contact for a whole string of cells has the smallest memory cell size as shown in Figure 2 In addition, when one adds with the capability of MLC (Multi Level Cells) to NAND devices, the bit density dramatically increases.

The NAND market has been continuously growing for the last several years. Figure 3 shows the NAND revenue and Gigabytes increase since 2008 and the forward projection for the years 2012-2016.

As the NAND technology has been moving to smaller and smaller process nodes some serious problems, physical and electrical surfaced:
           Physical Limitations:

• Pattern scaling - lack of EUV is a major issue
• Structure formation, Figure 4 depicts a 27nm NAND cell that shows how close the cells are getting to each other, and how much the aspect ratio is getting out of hand. This is a limiter to obtaining high yield.

          Electrical Limitations:

• There is an increase in cell-to-cell interference in the word lines.
• Capacitive coupling ratio has decreased
• Dielectric leakage has increased

The number of electrons on the floating gate has decreased dramatically so much so that a small fluctuation in the number on the floating gate can make a huge effect on the cell function. Figure 5 describes the scaling induced phenomenon.

It is a common understanding among the experts that the current NAND technology will not be able to be scaled down to the 10nm node.

The solution for this dilemma is the 3D NAND, which was initially proposed by Toshiba at the 2007 VLSI Symposium [3]. Toshiba unveiled its Bit Cost Scalable (BiCS) technology. BiCS makes use of a “punch-and plug” structure and charge trap memory films. Toshiba has fabricated a prototype 32-Gbit BiCS flash memory test array with a 16-layer memory cell using 60nm design rules, see Figure 6. Hynix, Samsung and Macronix have also come with their versions of the 3D NAND.

The following are the key advantages of the 3D NAND:

• With 3D NAND, scaling is no longer driven by lithography. The gate length is defined by deposition
  
• The key steps to 3D NAND are
  

                                        - Build a multitude of oxide/nitride or oxide/doped polysilicon stacked layers
                                - Fill the deep memory holes or trench slits. The top foreseeable challenges are ultra-high-aspect ratio (>40:1) conductor etch and dielectric etch with high etch selectivity to the hard mask

• 3D NAND is relatively straightforward for a DRAM maker since it has stacked SiO2 and polysilicon layers like a stacked capacitor DRAM, and trenches like a trench cell DRAM. 
  
• 3D NAND is evolutionary, not revolutionary. 
  
• The good news is continued cost reduction, smaller die sizes and more capacity. 
  
• Installed NAND toolsets in the wafer Fabs can, for the most part, be reused, thereby extending the useful life of Fab equipment. 
  
• 3D NAND technology is still basically NAND with all its inherent limitations of data reliability and performance: hence, generally well understood (evolutionary).
  

At this point all the NAND companies are putting a lot of effort to bring this process to high volume manufacturing; the current expectations are that in 2014-2015 it will be ready for prime time. 3D NAND will be a technology that will take us between the 2D planar NAND and whichever post-NAND technology emerges in the future.

Figure 7 describes the essence of the advantage of moving from 2D to 3D NAND. The adoption of 3D NAND technology will remove the burden from the Litho (and hence EUV) into the much easier process steps (deposition). Of course there are other advantages as described above.

It is not too difficult to see the similarity between the up and coming 3D NAND and the Monolithic 3D approach. As we describe in our web site (www.monolithic3d.com) the advanced technology patented by MonolithIC 3D Inc. enables the fabrication of Monolithic 3D Integrated Circuits with multiple stacked transistor layers and ultra-dense vertical connectivity. Thus, it appears monolithic 3D-ICs with 2 device layers provide benefits similar to a generation of conventional scaling. Furthermore, just as conventional scaling reduces feature sizes every generation, monolithic 3D opens the road for many years of continuous scaling by ‘folding’ once, twice, and so forth without necessarily reducing feature sizes.

1. F. Masuoka et. al IEDM 1984 pp464-467
2. F. Masuoka et. al IEDM 1987 pp552-555
3. H. Tanaka et al., Symp. on VLSI Tech. Dig., pp 14-15, 2007