Giacomo Gramaglia
HW/SW Co-Design of a Fault-Tolerant Accelerator for Sparse Transformer Self-Attention Through Packed Token Pruning.
Rel. Luciano Lavagno. 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 (5MB) | Preview |
Abstract
The transformer architecture has become dominant in fields such as natural language processing and computer vision, as its self-attention mechanism enables greater parallelism than convolutional and recurrent architectures. This allows significant speedups when run on parallel hardware, such as GPUs or tensor processors, and has driven a fast-paced growth in LLM sizes and context lengths. However, the high computational complexity of transformer self-attention, which grows quadratically with sequence length, combined with its high memory bandwidth requirements, poses a serious challenge to the architecture’s scalability. Memory usage is particularly critical, as memory band width has become the primary bottleneck in most AI workloads.
An additional challenge often overlooked is posed by hardware reliability
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
Ente in cotutela
Aziende collaboratrici
URI
![]() |
Modifica (riservato agli operatori) |
