MNEMOSENE: Computation-in-memory based on resistive devices
Named after the Greek goddess of memory, the MNEMOSENE project aims to demonstrate a new device that can combine the memory and computational parts of today’s computers. Such new computation-in-memory (CIM) devices can drastically increase both performance and energy-efficiency. Professor Said Hamdioui coordinates the 4 million euro project. “This can increase our computing efficiency immensely, at a thousand times less energy consumption”, he says.
In today’s computers, different parts deal with memory and computation. “There is continuous communication back and forth between the memory and the CPU to get data and process them”, explains Said Hamdioui, Professor of Dependable and Emerging Computer Technologies and head of the Computer Engineering Laboratory and of the Quantum and Computer Engineering Department. The energy consumption for communication and memory access can take up 80% of the total energy. “It is like having ten euro for shopping, and then having to spend eight euro on a taxi taking you to the shop, leaving you with just two euros to buy the essentials.” At the same time, demands on storage and computing power are ever increasing, with big-data applications and the internet-of-things (IoT) shaping our future. New applications require the kind of computing power that was the province of the supercomputers of yore. Moreover, that computing power should preferably come in small devices that consume as little power as possible.
Memristors are simple devices
To deliver that, we need new architectures, and MNEMOSENE is proposing just such an architecture. “What if, instead of continuously moving the data back and forth, we move the computing into the memory?” says Hamdioui. “Luckily, there are emerging devices that have the potential to perform both storage and computing activities, enabling such a new paradigm. We call them memristor devices.” Memristor is a contraction of the words memory and resistor; simply put it is a resistor with memory. Like resistors or capacitors, they are two-terminal elements. “It is a very simple device, compared to a transistors that has four-terminals. Memristive devices are also compatible with today’s semiconductor technology, so they can easily be mass-produced.”
Computation-in-memory is close to realisation
Still the new kid on the block, memristors are about to make a breakthrough. Hamdioui: “Several of them are being prototyped at the moment: Intel has been working on STT-MRAM, IBM on PCM, Micron on RRAM; all these devices can be seen as different flavours of memristors that are bringing computation-in-memory close to realisation. We deal with emerging devices, so there are many challenges to be solved. Nevertheless, demonstrations of computation-in-memory are underway. For example, together with IBM we are currently investigating the potential of performing bit-wise logic operations within the memory. This is extremely useful for applications such as hyper-dimensional computing, language recognition, bio signal processing or encryption.”
No need to keep charging
Part of the energy savings comes from eliminating the need to load data from the memory to the processor registers and storing the result back in the memory after performing the computation. But that is not all: “Memristive devices are non-volatile, meaning there is practically no leakage current, making for further energy savings. Unlike for example your smartphone that you have to keep charging all the time”, Hamdioui says. That also means that such devices retain their memory. “Memristors are very small; they could be manufactured even below ten nanometre. If the technology matures enough, you can integrate billions of them on a single piece of silicon, enabling huge storage facilities.” Though that is somewhat beyond the scope of this project. “CIM is a new paradigm. We don’t have CIM computers yet; CIM means changing programming models, the architecture, the way we do circuit design, and so on. This project aims at demonstrating the concept and its potential. We want to show that it works, at least at a small scale.”
Second time lucky
Originally, Hamdioui submitted his proposal for MNEMOSENE to the Future and Emerging Technologies (FET) programme. “We scored very high with that proposal, 4.8 out of 5. However, the acceptance rate is extremely low, less than five percent. We then reshaped it and resubmitted it to H2020, and it was accepted in the first time”, he remembers. “We have brought together a strong consortium, including IMEC who are one of the top in the world when it comes to research on cutting-edge semiconductor technology.” Other partners are ARM Holdings, a British multinational semiconductor and software design company, IBM Switzerland, TU Eindhoven, ETH Zurich, Aachen University, and INRIA, the French National Institute for Research in Computer Science and Automation.
Take advantage of the experience of others
It is far from his first time as coordinator: “Since I joined TU Delft, I have been involved in lots of projects, and coordinated many proposals”, he says. He was not always so successful as with MNEMOSENE. “When I just joined the university as an Assistant Professor, I lost so much time on writing proposals without any help. It is a huge effort with a terrible success rate. I have since learned how to optimise my efforts.” Hamdioui is in the process of preparing a few new proposals, and he is also guiding one of his colleagues. “Someone in the early stages of their career can use all the help they can get. With success rates getting even lower, you have to think about ways to increase your luck, and taking advantage of the experience of others is very important in my view. That will have a large impact on the quality of your proposal.”
You need contacts to succeed
He believes the university has a responsibility here. “We simply cannot hire people who have never learned how to play this game and give them a few years to find their own funding. We have to help them”, he says. “Let us create a network of people who have been successful that beginners can go to and talk about what to do and what to avoid. And let’s share our contacts with them. I was lucky in the sense that I had contacts from my years working in industry, but why would for example the IBM’s and NXP’s of this world work with you if they do not know you? I am certainly willing to offer that kind of help to all of my colleagues.”
Too much control is not a good thing
Coordinating the project itself is relatively easy, Hamdioui believes. “It is still a time investment with all the reporting. The rest is mostly managing people, and that is what I do on a daily basis”, he says. “You should try to inspire people, to provide them with the right circumstances, and give them freedom. Managing professionals too much is not a good thing; it has a negative impact on their productivity.” Initiating and coordinating a project is also a good way for young researchers to learn and to prove themselves: “If you have a good idea, then just develop it a little and lobby for it to get the right partners. Just go for it. Even if you fail, you will learn a lot. This will not only help with your visibility and speed up your development, but will also provide you with a great opportunity to interact with and learn from different researchers from both academia and industry. I strongly believe that the best way to collaborate with others is through these projects. Because then, whether you like it or not, you have to work together and deliver.”
'Let us create a network of people who have been successful that beginners can go to and talk about what to do and what to avoid. And let’s share our contacts with them'