If It's Tuesday This Must Be Belgium. My First Visit to imec

Outside of semiconductor, Belgium is famous for three things: beer, chocolate, and fries (which they eat with mayonnaise). Inside semiconductor, it is famous for imec. It's also famous for the 1969 movie If It's Tuesday, This Must Be Belgium. And it is, indeed, Tuesday.

After CDNLive in Munich (on Tuesday, to keep with tradition) I flew to Brussels and took the train to Leuven. After 35 years in the semiconductor business, it is hard to believe that it was the first time I visited. I've been to ITF, the Imec Technology Forum, a couple of times, but they hold that in downtown Brussels, not out at Leuven.

Getting There

In case you ever have to visit imec, here's the easy way to get there. Fly to Brussels. If you need to stay overnight, stay in the Sheraton which is literally across the street from top level of the airport terminal. There is a train station in the basement of the terminal. You can buy a ticket at the machines to Leuven (although I had to type it as Louvain, its French name). They run roughly every half hour. At Leuven, take a taxi from outside the station to imec. If they ask you where at imec you want to go, tell them "the tower" since that is where the visitor reception is.

Imec History

When I lived in France, imec existed. It was created in 1984. However, it was more European-focused than it is today, with its main partners being the three big European semiconductor companies:

• Siemens, who spun their semiconductor operations out to create Infineon and Qimonda (RIP)
• Philips, who spun their semiconductor operations out to create NXP
• ST Microelectronics (although, if you go back far enough, SGS and Thomson Semiconducteurs pre-merger)

Back in that era, most semiconductor companies had a research fab, often called the TD fab since the generic-sounding "technology development" has the specific meaning of developing process technology inside a semiconductor company. But as the cost of fabs, fab equipment, and the investment to create a process rose, there was a need for a research environment with a fab that could allow semiconductor companies and equipment companies, to cooperate in the pre-competitive time-frame. There can only really be one roadmap for the industry since the equipment and material suppliers have to line up behind it. That's not possible if each semiconductor company develops its own esoteric flavor or processes requiring different things from its suppliers.

Imec is in Belgium because that's where it started. But it is neutral ground, in the same way as it is a neutral country for the headquarters of the European Union. If it had been in the US, Taiwan, Japan (and later Korea or China) then it would not have worked. SEMATECH was a similar operation in its early days, except its mission was explicitly to regain US competitiveness after Japan had driven companies like Intel out of the memory business, and looked like it might dominate the entire semiconductor business. But even later when it threw its doors open to non-US companies, it was still in the US and so not seen as independent.

Today, all the leading edge semiconductor companies in both logic and memory have joint programs with imec. They each send a couple of dozen engineers, over 500 "guest" engineers in total, to work on programs there. Cadence has a couple of people there, as do many of the equipment vendors, and other contributors to the ecosystem.

My Visit

I met first with Luca Matti, who somehow has three hats, working at imec, while finishing his Ph.D. at Braunschweig University of Technology, and being part of the Cadence Academic Network. His work is on DTCO, Design-Technology Co-Optimization. He was also involved in the recently announced Cadence-imec 3nm testchip. See my post 3nm Cadence and imec for what I wrote at the time we announced it. Luca and Diederik Verkest gave me a bit more detail. The chip was not a whole chip, just a large block containing an Arm processor. However, the whole block was laid out and went through timing closure. That way, every aspect of the block was realistic. However, only a few masks were made since the main purpose was to investigate manufacturability of the tightest layers.

Diederik gave me imec's view on what the future roadmap for the semiconductor industry will be. Since the leading semiconductor companies who all have staff at imec, this is really the consensus roadmap. I will cover that in more detail in a post of its own.

Kurt Ronse gave me a detailed look at what the remaining issues with EUV are. I'll cover that in another post too.

I was taken on a fab "window tour". That means I got to see into the fab from outside, but I didn't suit up in a bunny suit and go inside. They have just extended the fab with an expansion area which is now being filled with equipment. The total investment is over €1B and I suspect that mostly covers the shell, air conditioning, water conditioning, and so on. Most of the equipment is also part of research programs.

The most interesting piece of equipment was, of course, "the beast", the ASML 3300 EUV stepper. It was especially interesting to see since it was undergoing maintenance and so most of the covers were off revealing the innards. I'd love to show you a picture of it, but it was too interesting—I wasn't allowed to take photos of that part of the fab. But I was allowed to take some in less sensitive areas of the fab.

Memory

Arnaud Furnemont gave me a presentation about where imec sees the future of memory. His summary was "it's all 3D NAND and DRAM". That's for standalone memories. For embedded memories, some other technologies are attractive. Arnaud said that a few years ago it looked like RRAM was going to take over the world, with performance close to DRAM and non-volatility. But the current required to make the filaments stable seem to be too large and so we have three memory technologies for standalone:

• 3D NAND. The number of layers will continue to increase, but nothing dramatic is expected to change.
• Storage class memory, which is the generic name for Intel/Micro 3D Xpoint. But that has been slow in really getting to market, and there are no other suppliers of similar technology.
• DRAM. It will continue to improve slowly at about 1.1X per year. Adding on-chip ECC is helping since it makes it easier to manage the fact that some of the transistors on a memory die will be a long way from the mean since there are billions of them.

The big memory challenge is that demand for memory, especially from mobile and IoT, is increasing faster than the memory technology. Looking at the big picture, that means that the only way to scale fast enough to keep up is to build more fabs. China is building lots; in particular, a NAND flash, a NOR flash, and a couple of DRAM fabs. But even looking at the entire world, we can't get onto a sort of Moore's Law of fabs where we have to build twice as many memory fabs each year to satisfy demand. But there don't seem to be any breakthrough technologies on the horizon even in research, which means nothing in volume manufacturing for five to ten years.

In embedded memories, MRAM is the most promising technology, although RRAM also has a place. eFlash is too expensive in terms of masks, so both MRAM and RRAM are attractive. They don't involve so many masks, and they live in the BEOL so it is easy to move from one transistor-level technology to another. This fits with what the foundries seem to have on their roadmaps.

The Beast

Here's a picture of the EUV stepper, to give you an idea of how large it is. I didn't take this, it was taken by imec's official photographer. But it is far and away the largest piece of equipment in the fab, and with all the other stuff in the way, impossible to get into a single frame. This is not actually the 3300, but the prior generation 3100, but I'm sure you already noticed that!

Sign up for Sunday Brunch, the weekly Breakfast Bytes email.