Boosting computer power

What is the advantage of having the function of processor and storage combined in one unit?

Gambardella: Currently we are running into a memory wall. Our computer applications process more and more data and we keep shuffling the data from the memory to the processor and back. At some point, it does not matter anymore how fast your processor is because you cannot keep up with moving the data to-and-fro, let alone the energy consumption that goes along with the data movement. We want to reduce that data movement.

By embedding the logic into the memory unit?

Hrabec: Exactly. In a joint collaboration of ETH Zurich and the Paul Scherrer Institute, we built a prototype of a logic-in-memory device. Our collaborator Zhaochu Luo proposed a rather bold design for a NAND gate, which is one of the building blocks in logical circuits. We were thrilled to see it work on the very first attempt. Then, nothing could stop us anymore and we built a functional full adder made of 15 NAND gates. The realisation of the NAND and full adder gates demonstrates that our circuits are functionally complete and that several gates can be cascaded with one another, which is a main requisite in electronic logic devices. 

What is the difference to a logic gate built from transistors?

Gambardella: When you power off a transistor it forgets its digital state. Our logic gates do not need electricity to remember their inputs and outputs. The data is stored magnetically. Consequently, our logic gates are always instant-on.

Can you tell us more about what your logic in-memory device looks like?

Hrabec: Our data is stored along segments on nanowires (see fig. 2). The segments comprise magnetic domains with two different magnetic orientations representing zeros and ones of binary code. We can push the magnetic domains along the wires and through the logic gates by applying a current. A special type of spin coupling inside the gates determines the magnetic orientation of the output. The domains move at a speed of about 100 metres per second. If you translate this speed to the size of our tiny NAND gate, you reach the clock speed of modern-day CPUs. In fact, we think it is possible to speed up the magnetic domains further – maybe up to one kilometer per second. This would outpace the clock speed of any existing CPU.

What step needs to be taken towards a successful market implementation?

Gambardella: This is an important question. We need complementary expertise from outside our groups to complete the integration of logic-in-memory devices. We need a read/write head based on a so-called magnetic tunnel junction. The head has to be small enough to access the tiny magnetic domains to read the stored magnetic spin and to manipulate the magnetic spin for writing data.  We believe that big companies like IBM or Samsung or dedicated technology incubators could perform this task. We are actively seeking collaboration partners to develop further this technology.