Security and Resilience of Non-Volatile Memories Based on In-Memory Computing
Open Access
- Author:
- Sayyah Ensan, Sina
- Graduate Program:
- Electrical Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- March 14, 2022
- Committee Members:
- Swaroop Ghosh, Chair & Dissertation Advisor
Joseph Najem, Outside Unit & Field Member
Abhronil Sengupta, Major Field Member
Morteza Kayyalha, Major Field Member
Kultegin Aydin, Program Head/Chair - Keywords:
- In-Memory Computing
Resiliency
Security
RRAM
Floating Point - Abstract:
- Von-Neumann computing has separated computing and storage units, and this leads to incapability to meet the challenges introduced in big data era. Transistor scaling has offered faster processing and storage units over years but the transfer bus between these two units (Von-Neumann bottleneck) is dominating power, energy consumption, and performance of the Von-Neumann computing model. In-Memory Computing (IMC) is one of the most promising architectures that can solve this challenge locally. The basic idea of IMC is to infuse computing abilities into the storage elements. IMC is achievable by using Non-Volatile Memories (NVMs) such as Spin-Transfer Torque RAM (STTRAM), Magnetic RAM (MRAM), Phase Change Memory (PCM), and Resistive RAM (RRAM). RRAM characteristics such as low power consumption, fast operation, and high integration density (4f^2 footprint in crossbar architecture) make it a suitable candidate for NVM application. However, on the other hand, RRAM characteristics such as high and asymmetric read/write current, retention time, and defects during fabrication bring new threats to security and resiliency of the IMC circuits. IMC circuits will be a huge part of future chips and are believed to become ubiquitous in future computing devices. Therefore, it is very important to investigate their security against adversaries and resiliency against defects. Note that, the challenges for security and resiliency of compute-capable NVMs is completely different from their storage counterpart. In this work, we have implemented two IMC systems which first one is capable of performing Floating Point (FP) arithmetic operations and the other one is a first order linear Partial Differential Equation (PDE) solver. Later on in this work, we investigate the possible adversary attacks against two general purpose IMC architectures namely Memristor Aided Logic (MAGIC) and Dynamic Computing In-Memory (DCIM) and then we propose countermeasure approaches to obfuscate the implemented function from the adversary. Furthermore, we propose approaches to meet the challenges arising from defects in RRAMs’ fabrication. These approaches make sure that IMC architecture can work if there are some RRAMs Stuck At Fault (SAF).