Georgia Institute of Technology
Compute-in-memory for Efficient AI: Cross-layer Modeling and Hardware-software Co-Design
Abstract
dc:description.abstractModern deep learning is bottlenecked by the memory wall, where moving weights and activations consumes several times the energy of doing computations. Compute-in-memory (CIM) is one of the main ways around the wall; by performing multiply-accumulate operations inside the memory arrays that store the weights, analog CIM designs project efficiency gains of one to two orders of magnitude over GPUs. Those gains, however, are often reported with insufficient analysis of the impact of analog noise on the accuracy of the model. In real-world deployment, a tolerance for model accuracy comes first, and efficiency is then maximized within that tolerance. The efficiency of a CIM design therefore has to be evaluated jointly with the accuracy it delivers, and building one is a hardware-software co-design problem: full-model accuracy must be co-optimized with energy, area, and throughput.
Degree
thesis:*- Name thesis:degree_name
- Electrical and Computer Engineering, PhD
- Department thesis:degree_department
- Electrical and Computer Engineering
- Grantor
- Georgia Institute of Technology
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Read, James
- Advisor dc:contributor.advisor
-
- Yu, Shimeng
- Committee members dc:contributor.committeemember
-
- Khan, Asif
- Gavrilovska Habl, Ada
- Lim, Sung-Kyu
- Hao, Callie
- Bakir, Muhannad
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/82224