3.6 Hot Topic - Memristor based Computation-in-Memory Architecture for Data-Intensive Applications

Date: Tuesday 10 March 2015
Time: 14:30 - 16:00
Location / Room: Bayard

Organisers:
Said Hamdioui, TU Delft, NL
Koen Bertels, TU Delft, NL

In today's data-intensive applications (known as Big Data problems), such as healthcare (e.g., use of genetic information to diagnose and treat diseases), social media, engineering (e.g. large scientific experiments), the primary goal is to increase the understanding of processes in order to extract so much potential and highly useful information hidden in the huge volume of data, which in turn can be used to increase the productivity. As the speed of information growth exceeds Moore's Law at the beginning of this century, excessive data is making great troubles to human beings. At the same time, Big Data arises with many challenges, such as data capture, data storage, data analysis and data visualization. Performing data analysis within economically affordable time and energy is the pillar to solve big data problems, and therefore extract extremely valuable information. The increase of the data size has already surpassed the capabilities of today's computation architectures which suffer from communication bottleneck due to limited bandwidth. For instance, the transfer of 1 petabyes data at a rate of 1000MB/second will cost 12.5 days! Communication and memory access does not only kill the performance, but also energy/power (more than between 70% and 90% such applications). Even the CMOS technology used to implement today's architectures contributes to such power due to the higher leakage; not to mention the limited scalability (as it is becoming very costly), reduced reliability (as it degrades faster), etc. In conclusion, today's CMOS based architecture are not able to provide the computation capability needed for data-intensive applications. New architectures based new technologies are therefore needed. This Hot-Topic Session will address the concept of "Computing-in-memory (CIM)" and discuss a new Memristor Based Architecture Paradigm for Data-Intensive applications, as an alternative architecture. The concept is based on performing the storage and computation in the same crossbar topology (non Von- Neumann architecture) where the key device is the non-volatile resistive switching element (memristor). CIM architecture is able significantly push the "memory wall", while the memristor device is able to reduce the static power to practically zero.