Place : auditorium Z704, Phelma, 3 Parvis Louis Néel, 38000 Grenoble
video conference : https://visio.numerique.gouv.fr/guk-owrj-fwv
Abstract : The growing demand for deploying deep neural networks on edge devices is constrained by the energy cost of moving data between memory and processing units. Analog in-memory computing (AiMC) addresses this limitation by performing matrix-vector operations directly within resistive memory crossbars. Among emerging memory technologies, spinorbit torque magnetic RAM (SOT-MRAM) offers high endurance, CMOS compatibility, non-volatility, and separated read and write paths, making it an attractive candidate for AiMC accelerators. This thesis first develops a scalable SOT-MRAM architecture for binarized neural networks using high-resistance 4T2R XNOR bitcells and time-domain column sensing. The accumulated current is converted into time and digitized with a time-to-digital converter, avoiding conventional expensive ADCs. Circuit-level Monte Carlo simulations, calibration, and hardware-aware fine-tuning show that the sensing precision can be adapted to the target workload to improve energy efficiency and throughput with limited accuracy loss. Binary computation, however, achieves efficiency at the cost of numerical precision. Higher-precision weights generally require bit-slicing across multiple binary cells or arrays, increasing array and peripheral overhead. To address this limitation, the thesis investigates multi-pillar voltage-gated SOT (VGSOT) MRAM cells for native multilevel storage. Multiple MTJ pillars share a SOT track and can be selectively programmed through voltage-controlled magnetic anisotropy, allowing several resistance levels to be represented within one cell structure. A corresponding multilevel AiMC architecture is developed using the same time-domain sensing principle. The binary and multilevel approaches are then compared through a common systemlevel framework that accounts for crossbar computation, digital processing, and memory traffic. The CIFAR-10 energy breakdown shows that binary crossbars perform most weighted operations, while the arrays, their TDC readout, and dedicated digital hardware together account for 0.7 to 1.7 % of the inference energy. Full-precision layers executed on the CPU account for 90 to 98 %. In the corresponding multilevel implementations, all weighted operations are mapped to the arrays, and the arrays with their TDC readout account for 40 to 77 % of the energy. The best multilevel variants improve accuracy by 2.05 percentage points on CIFAR-10 and 5.10 percentage points on CIFAR-100. Overall, the thesis shows that the cost of numerical precision in AiMC emerges from the interaction between the memory representation, sensing architecture, network mapping, and surrounding digital system. It establishes a device-to-system methodology for evaluating MRAM-based AiMC and highlights multi-pillar VGSOT-MRAM as a promising path toward higher-precision computation without the replication imposed by binary bit-slicing.
Jury :
- Marc Bocquet, PROFESSEUR DES UNIVERSITES, IM2NP, Aix-Marseille Université, Rapporteur
- Antoine Frappé, PROFESSEUR DES UNIVERSITES, IEMN, Université de Lille Rapporteur
- Ian O’Connor, PROFESSEUR DES UNIVERSITES, Ecole Centrale de Lyon, Université de Lyon, Examinateur
- Katell Morin-Allory, ASSOCIATE PROFESSOR, TIMA, Grenoble INP, Examinatrice
Thesis supervisors :
- Guillaume PRENAT, DIRECTEUR DE RECHERCHE, CEA CENTRE DE GRENOBLE, Directeur de thèse
- Pascal BENOIT,Professeur, Université de Montpellier,Co-directeur de thèse
- Jonathan MIQUEL,Maître de conférences, Grenoble INP – UGA,Co-encadrant de thèse
- David NOVO,Directeur Recherche, CNRS,Co-encadrant de thèse
