An Steegen, Executive Vice President Semiconductor Technology and Systems at imec
The zettabyte era
In this digital age, we consume a few zettabytes (= 1012 GB) of data each year. Predictions are that by 2025 this will increase to a few hundred zettabytes/year. And this comes at a price: to store one zettabyte you need 1,000 datacenters with each datacenter consuming the equivalent of power of 180.000 homes.
But there’s not only a challenge related to energy consumption. Today’s processor chips – the ones in supercomputers – can perform Terabyte operations per second, or 103 GB per second. We have optical links transferring 3 TB per second. And state-of-the-art memories can store up to 1 TB. So we still have a long way to go if we want to efficiently handle ZB of data. Partly, we can use conventional scaling to build better performing processors, networks and memories, but it is clear that we will need to turn to smarter ways to compute, store and connect.
Storing data in a molecule
Together with industry leaders we develop future memory technologies. Our innovations cover high speed embedded on-chip cache memory such as SRAM & STT-MRAM, scaled high speed dynamic random access memory (DRAM) devices, new storage-class memories for massive data access in short time, such as RRAM and improved NAND Flash memory devices for high performance and mobile devices. But we also recently launched a fascinating new memory research topic.
Just like hard disks store digital information, DNA molecules store biological information, within each person’s cells. Would it be possible to store digital data in DNA molecules? In a Harvard experiment, researchers succeeded in storing 50,000 words in a DNA molecule. The big advantage of DNA storage is that it is extremely dense and has a long lifetime. You could store petabytes (= 106 GB) of data in one gram of DNA, for about 500,000 years.
With the current techniques, storing data in DNA is a slow and expensive process. At imec, we have many expertises under one roof, including semiconductor and memory technology, biotechnology, biology, chemistry. Cross-linking these expertises will be key to build an affordable, manufacturable and massively parallel silicon DNA storage platform.
The interaction between these various experts also works in the other way – using advanced logic devices such as FinFETs and nanowires – to read DNA. Indeed, new and innovative ways to read DNA are needed to make DNA sequencing faster, cheaper and simpler as it will more and more evolve towards a mainstream diagnostic tool.
Turning data into wisdom
Turning all the zettabytes of data into knowledge is one of the major challenges of our time. As mentioned before, we will first turn to the conventional scaling techniques to better compute, store and connect. As industry is now implementing the leading-edge 10nm logic technology, research hubs such as imec are developing the next nodes – 3nm and beyond – by building new architectures, using new materials etc. We still succeed in achieving more speed, and in maintaining the power density. This road along Moore’s Law will go on for still some years.
And in the same way, we are leveraging new memories to distribute power and performance more efficiently in the system. Also, 3D packaging and optical links become extremely important to manage the power and performance in the hybrid technology platforms of the future.
And although it will require major efforts and innovative power, we hope to get a 2x speed and 2x density for every generation, using the conventional scaling concept. But this won’t last and will not be enough to handle all the data and to get meaning out of it. How than will we create more ‘wisdom’? How will we make a machine, like a computer, to actually do something with the data that it has collected?
The solution bears different names: deep learning, machine learning, neuromorphic computing, or the global name: artificial intelligence. It’s the new buzz word and imec is also active in this field. Thanks to the merger with iMinds, we now have several imec groups with a lot of experience in machine learning techniques. It will be important to select the right machine learning technique for the right application.
Making quantum computing a reality
There are some problems that can’t be solved with any regular or conventional compute engine that we have today. One example is the extremely complex structure of proteins, and the protein folding process. Capturing this is important for understanding – and curing – many diseases.
To unlock this knowledge, all hopes are set on quantum computing. Simply put, it’s a way of computing not only using 0’s and 1’s, but also all the fuzzy states in between. It tackles problems from every side and in a massive parallel way. It doesn’t come up with one solution but with 1000 nearby solutions. Problems like protein folding will in the future be tackled by quantum computing. Until now, mostly academic groups were studying quantum computing building blocks. Imec, with its expertise and infrastructure, and with its scaled technology platforms, launched a new activity, aiming to build a manufacturable and scalable quantum technology platform, with a system-level approach.
5G connectivity based on scaled technology and innovative circuit design
5G networks are believed to become commercially available from 2020 on. It’s the next-generation of mobile networks beyond the 4G LTE mobile networks of today. These networks will require new, more complex front-end modules and more advanced integration schemes, linking the antennas to the RF chip to the modem. Scaled technologies will be an indispensable building block in all this, as well as high-speed analog RF devices.
Imec developed a lot of innovative solutions at the circuit level, using off the shelf foundry technologies. This will now be combined with our logic and process knowhow in III-V to build super-fast high-speed analog devices. This will bring a whole new level of innovation in 5G components, exploiting the real benefits of scaled technology and couple this with very advanced circuit design. We’re doing this for very high speed applications, but the same can be done for very low-power applications. A similar approach, crosslinking technology and system knowhow, are we applying to mm-wave technologies and sensing solutions.
An overview of the smart society building blocks
As shown in the picture below, there are different levels of innovation needed to realize the smart society. The smart society is asking for new applications, which will rely on a new kind of infrastructure to compute, store and connect the immense amount of data. Next to this, new devices, sensors, actuators etc. will need to be developed, and system functions will need to be improved. Technology platforms and innovations in process steps are at the base of all these new applications. Imec is uniquely placed to create this future.