Michele Solimine
Design-Space Exploration of Memory Systems for Precision-Scalable Systolic-Arrays.
Rel. Mario Roberto Casu, Edward Manca. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Elettronica (Electronic Engineering), 2026
|
Preview |
PDF (Tesi_di_laurea)
- Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (2MB) | Preview |
Abstract
Recent advances in artificial intelligence and deep learning are driving the demand for hardware architectures that are efficient performing the inference of Deep Neural Networks (DNNs). At the core of these algorithms we have Matrix-Multiplications (Mat-Muls). Systolic Arrays (SAs) are one of the most used architectures to compute Mat-Muls. They are commonly integrated into state-of-the-art (SOTA) Application-Specific-Integrated-Circuit (ASIC) such as, e.g., the Google TPU. A SA is a 2D matrix of Processing Elements (PEs) with neighbor-to-neighbor (N2N) connectivity. This matches the data reuse of the Mat-Mul algorithm, and simplifies the global routing in silicon implementation, achieving better Power-Performance-Area (PPA) metric. Most of the time, DNNs inference is performed using reduce precision integer data format, e.g., from 8 to 2-bit integer operands.
Recently, the PEs of SAs have been improved by integrating Precision-Scalable (PS) multipliers, that supports many integer precisions and can increase the number of operations performed when operands precision is reduced
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
URI
![]() |
Modifica (riservato agli operatori) |
