Tag Archives: IBM

Can we take cost out of technology scaling?

There is much talk these days about continued scaling, including some recent posts by my colleague Ed Korczynski, in “Moore’s Law is Dead” Part 1 (What?) and Part 2 (When?). At The ConFab in June, keynote speaker, Dr. Gary Patton, vice president, semiconductor research and development center at IBM, talked about scaling, adding some historical perspective. I previously blogged about the “three fundamental shifts” that Patton believes will lead to a bright future for the semiconductor industry.

“We will keep scaling,” he said. “We have shown a tremendous ability to innovate and keep moving that technology forward.”

In the 1990s, Patton notes that life was actually pretty simple. “You brought in a new lithography tool, you scaled the horizontal dimensions, you scaled the vertical dimensions and you got a new technology out. It was better performance, the same power density, and you could do a lot more on the chip,” he said.

Patton_Slide5

Around 2000, we hit the gate oxide limit. “Gate oxide got to be abount three atomic layers. We could have said at that point ‘game over, scaling has ended.’ But guess what, we innovated. We came up with a pretty fundamental shift in ideas which is let’s change the fundamental properties of silicon. If we can strain the silicon, we can enhance the mobility. We can change the gate oxide. We can enhance the coupling between the gate and the channel. And that’s what we get over that last decade. We said let’s go from SRAM to a very high performance eDRAM (embedded DRAM) so we can put a lot more memory next to the processor because we knew memory was a key gating factor for the processor speed. This enabled the personal computing era and smart consumer electronics,” Patton said.

In 2010, we were at another one of these inflection points. “It’s not surprising that the improvements in 20nm are less than people would like because we really reached the end of the planar device era. Again, we were saying ‘game over, we’re done scaling.’ But no, we continue to innovate. The next decade is really about 3D. 3D devices, finFETs, or 3D chip integration,” he added.

Patton said that design technology co-optimization will be a key piece of getting through the next decade. “That will probably take us to about 2020,” he said. At that point we’re going to “hit the atomic dimension limit and we’re going to have to do it all over again. Here, we’re going to get into nanotechnology. Nanowire devices, silicon nanowires, carbon nanotubes, photonics and multi-chip stacking to bring things together. That will enable wearable computing, everywhere connectivity and cognitive computing.”

Patton said the problem is not physics. “We’re going to have solved the physics problems,” he said. “The problem is financial.” Patton showed a chart (below) that depicted the history of our industry from 1980 to present. “What drove the industry was smaller features, which enabled better performance and better cost per function. It enabled new types of applications, and that enabled larger markets. If you look in this time period, there’s about a six order of magnitude improvement in cost per transistor and that enabled a seven order of magnitude increase in consumption of silicon transistors,” he said.

Patton_Slide6

The challenge we’re facing right now is depicted below, showing the compound growth rate reduction and the cost of a circuit. On the x-axis is linear scaling. “We’ve typically targeted about a 0.7X linear scaling, which means from an area perspective, you get about 50% improvement. Note the line, 50%, doesn’t go through 50% improvement because with each new technology, there is some increase in complexity. It might be more like 30% improvement at the die level. If we’re really good and provide some enhancements in the technology, self-aligned processes, things like that, we may get it to 40%. So 30-40% is about the range we’ve been getting in terms of the cost per die improvement as we scale up.

Patton_Slide7

“The challenge we’re facing now is two fold. Number one, we’re struggling to get that 0.7X linear scaling. It might be about 0.8X. And we’re adding a lot more complexity, especially when you adding double and triple patterning .The focus today in innovation has got to be heavily focused on ‘how do we drive cost?’ Not just how do you scale, because scaling would add a lot of extra cost at this point. How do we drive cost down, how do we keep adding value to the technology. The model is changing. Moore’s Law can still hold, but we have to focus on the cost equation. So there’s really two parts. Technology innovation which is focusing on the patterning, focusing on the materials, the processing, and how do we drive that to take cost out of the technology scaling,” Patton said.

Three fundamental shifts

At The ConFab last week, Dr. Gary Patton, vice president, semiconductor research and development center at IBM, said there is a bright future in microelectronics (I heartily agree). He said that although there seems to be a fair amount of doom and gloom that scaling is ending and Moore’s Law is over, he is very positive. “There are three huge fundamental shifts that are going to drive our industry forward, will drive revenue growth and will force us to keep innovating to enable new opportunities,” he said.

The first fundamental shift is the explosion of applications in the consumer and mobile space. Patton noted examples such as cars that can drive themselves and can detect people and bicyclists and avoid them, smart phones for as little as $25, wearable devices that not only tell you what you’re doing but how you’re doing, and 4K television. “That is an incredible TV system, but it’s going to demand a lot of bandwidth; twice the bandwidth that’s out there today. If you turn on your 4K system, your neighbors are going to start to notice it when they try to access the internet,” he said.

Patton said that it’s estimated that today there are about 12.5 billion devices connected to the internet. That’s expected to grow to $30 billion by 2020. This represents the second fundamental shift commonly known as Big Data. “All these interconnected devices are shoving tremendous amount of data up into the cloud at the rate of 1.5 Exabytes (1018) bytes of data per month,” Patton said. “And that’s grown by about an order of magnitude in just the last 13 years. The estimate is that in the next 4 years, it’s going to go up another order of magnitude. It’s accelerating.”

The third fundamental shift is with all this data going up into the cloud, the data is almost all unstructured data, such as video and audio. “It’s related data but disconnected. How do we take that data and do something with it? That brings us to analytics and cognitive computing. We have really just started in this arena.”

So there you have it. Three reasons to be very positive about the future of the semiconductor industry: an explosion of applications, the rise of big data and the need to analyze all that data.

IITC: New Materials for Advanced Interconnects

On-chip interconnects have not been scaling at the same speed as transistors. When TSMC went from 20nm to 16/14nm, for example, they decided to replace the bulk MOSFET with a FinFET, but they left the interconnect stack as is. In part, interconnect scaling has been slow because companies don’t want to make too many major changes at the same time and introduce risk. Costs, of course, are also an issue. “When you’ve got ten layers of metal and let’s say six layers of those are close to minimum pitch, it gets very expensive once you start doing double patterning,” said Dr. Deepak Chandra Sekar, general co-chair of the upcoming 2014 IITC/AMC joint conference. “With the interconnect layers, people want to save litho costs. That’s one reason they are not scaling as much as they used to.”

But the major reason is that it’s difficult to make interconnects much smaller without introducing significant increases in resistivity. “If you scale down and your resistivity goes up exponentially, it can be a problem,” Sekar said. “Copper resistivity shoots up when you scale it down because of surface scattering, grain boundary scattering and interface roughness.”

The 17th annual International Interconnect Technology Conference (IITC) will be held May 21 – 23, 2014 in conjunction with the 31st Advanced Metallization Conference (AMC) at the Doubletree Hotel in San Jose, California. It will be preceded by a day-long workshop on “Manufacturing of Interconnect Technologies: Where are we now and where do we go from here?” on Tuesday, May 20.

Sekar highlighted a number of papers that will be presented this year. Many of them focus on new materials that could lead to reduced resistivity and enable further interconnect scaling. “There is a lot of excitement about carbon and carbon-copper composites eventually replacing copper,” he said. “At IITC this year, we have a couple of papers, one on graphene showing lower resistivity than copper, and then one on carbon nanotubes showing good resistivity as well. They are still a bit far out in the sense that there’s a lot more process integration work that needs to be done because these are proof of concept demos, but they show that there might be more beyond copper.”

In a paper from AIST, titled “Sub 10nm wide intercalated multi-layer graphene interconnects with low resistivity,” work will be presented that demonstrates 8nm wide 6.4nm thick graphene interconnects with a resistivity of 3.2uohm-cm, which is significantly better than copper with similar dimensions. This milestone for graphene interconnect research is expected to motivate the process integration research that is required to take the technology to the next level.

8nm wide graphene interconnects

8nm wide graphene interconnects

Carbon nanotubes (CNTs) have been explored as a material for vertical interconnects for many years since they can handle higher current densities than copper and offer ballistic transport. A paper from imec titled “Electron Mean Free Path for CNT in Vertical Interconnects Approaches Copper,” work will be presented that demonstrates a 5x improvement in electron mean free path for CNTs compared to previous work. The CNT mean free path of 24-74nm approaches copper. Contact resistance is improved significantly compared to previous work as well.

Carbon Nanotube (CNT) vias in integrated structures

Carbon Nanotube (CNT) vias in integrated structures

Another challenge to scaling of interconnects: reliability. Both time-dependent-dielectric-breakdown (TDDB) and electromigration lifetimes for interconnects drop rapidly when scaled. In work to be presented at IITC/AMC, IBM and Applied Materials will present a multi-layer SiN cap process is developed that shows higher breakdown and lower leakage compared to conventional SiCNH caps. Selective cobalt caps in combination with the multi-layer SiN cap are shown to provide a 10x improvement in electromigration lifetimes. Wrap-around cobalt liners in combination with the cap layer schemes are shown to provide a 1000x improvement in electromigration lifetimes. The paper is titled “Advanced Metal and Dielectric Barrier Cap Films for Cu Low k Interconnects.”

10x improvement in electromigration lifetimes with multi-layer SiN and selective cobalt cap layers. 1000x improvement in electromigration lifetimes with multi-layer SiN cap, cobalt cap and wrap-around cobalt liners.

10x improvement in electromigration lifetimes with multi-layer SiN and selective cobalt cap layers. 1000x improvement in electromigration lifetimes with multi-layer SiN cap, cobalt cap and wrap-around cobalt liners.

Of course, an alternative to making everything smaller by scaling is to go 3D. That will be addressed by a variety of papers, including one from CEA-Leti focused on 3D monolithic integration. While most of today’s through-silicon vias (TSVs) are in the 5µm range, monolithic 3D technologies offer TSVs in the 50nm range, which allows dense connectivity between different layers in a 3D-IC. In the Leti paper, such dense connectivity is shown to provide 55% area reduction and 47% energy-delay product improvement for a 14nm FPGA design. Transistor technologies that allow monolithic 3D integration are experimentally demonstrated. “When you make the TSVs smaller and smaller, you can reduce the length of on-chip wires as well by taking what’s on a single now and stacking them into two layers,” Sekar said. “That might save a lot of power and area. There’s been a lot of talk about monolithic 3D, but these are some of the first few experimental demonstrations showing that it’s possible.”

Monolithic 3D-ICs

Monolithic 3D-ICs

No technical barriers seen for 450mm

Paul Farrar, general manager of the G450C consortium, said early work has demonstrated good results and that he sees no real barriers to implementing 450mm wafers from a technical standpoint. Speaking at the SEMI ISS meeting in January, Farrar showed impressive results from, etch, CVD, PVD, CMP, furnaces, electroplating, wet cleans and lithography processes and said the inspection/metrology tools were in place to measure results. “I don’t believe we will find fundamental technology limiters,” he said. “But we will have to keep working to find ways to maximize the efficiency.” Gaining such efficiencies are critical in order to meet the cost-saving goals of the program. “In the end, if this isn’t cheaper, no one is going to do it,” he said.

G450C is a consortium based at the CNSE campus in Albany, NY. It is financed by Intel, TSMC, Samsung, IBM, GLOBALFOUNDRIES, and New York State (CNSE). “Our job is to make it as easy as possible to innovation and be collaborative between the semiconductor makers and our key friends in the industry who enable the 450 work to be done in an economic way,” Farrar said.

At the end of 2013, G450C at 34 tools delivered to its 50,000ft2 fab in Albany, with another 7 tools in place at partner’s facilities. “The FOUPS are going, the overhead transport is well underway and some of the cleanroom is actually starting to look like a cleanroom,” Farrar said.

Farrar started with etch results, saying they were “starting to see some pretty good data – 3 sigma at about 2%. Yes, there’s still some work to get to the very edge of the wafer but relatively good progress and good jobs on gas delivery, etc.

FarrarISS_F6

He showed good results with both oxide and silicon nitride CVD, with close to 1.5mm edge exclusion. “It’s very representation data from early in the program,” Farrar said, noting that they were starting to pattern some of the more complex oxides.

FarrarISS_F7

He said the goal for PVD was to demonstrate better than 5% uniformity. “We know we have step coverage challenges for both the 10 and 7nm nodes. There’s tremendous work going on in the injection rings for gases, high density plasmas from multiple RF sources, but again some progress to me made but pretty good data for right out of the chute,” he said.

FarrarISS_F8

CMP results demonstrated repeatability less than 4%. “Very good job done by our suppliers,” Farrar said.

FarrarISS_F9

Farrar described data from furnaces as reasonably good. “We still need to do more characterization at what I call the micro level,” he said. “We see some hot spots on the edge, but we’re starting to work on those.”

FarrarISS_F10

Also “pretty good data” from electrochemical plating (ECP) of copper. “Well done here,” Farrar said. “The challenge is thermal and pattern loading effects, and gap fill.”

FarrarISS_F11

More of the same with wet cleans. “We’re starting to see some pretty good particle data. We’re cleaning wafers relatively well. We are seeing a few things like what I would call micro-metallic contamination that can grow some things so we’re still working on that. But from a particle removal standpoint, pretty good unit process work,” Farrar said.

FarrarISS_F12

Farrar acknowledged that lithography remained as one of the biggest challenges in the 450mm transition, but showed good results from directed self assembly across a 450mm wafer, and said the consortium had a very strong partnership with Nikon. “We’re working with them and we’ve seen some tremendous progress at their factory,” he said. “I’m fully confident that we’ll have capability by July to run patterned wafers. Immersion is going to be the workhorse. I think that’s a key enabler to get to 450mm.” He said the industry would have to see how the economics of EUV played out later in time. “I don’t think it’s going to be early in time,” he said.

FarrarISS_F13FarrarISS_F14

Farrar seemed to draw hope from the earlier transition from 200mm to 300mm wafers, which started around 1998.  “By 2008, we were getting more than 2X the number of wafers per tool out compared to what was going in 2003. There was about a 70% improvement over 5 years,” he said.