Gianmarco Fortunelli
Design of a configurable multi-datatype arithmetic unit for ML workloads.
Rel. Guido Masera. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Elettronica (Electronic Engineering), 2026
|
|
PDF (Tesi_di_laurea)
- Tesi
Accesso limitato a: Solo utenti staff fino al 2 Ottobre 2029 (data di embargo). Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (1MB) |
Abstract
As deep-learning models continue to impose growing computational and memory demands, numerical precision has become a central architectural constraint. Narrower representations reduce storage, memory traffic, and arithmetic cost, but can constrain numerical range or precision and thereby affect model accuracy. Microscaling (MX) formats address this trade-off by associating a shared scale with a block of narrow floating-point or integer private elements, extending the block’s effective range. Realizing MX efficiently requires conversion, scale generation and packing. This is particularly relevant when activations are produced in standard floating-point formats, while weights are quantized offline and stored in MX format. Since weights and activations are quantized independently and may favor different private element formats, forcing both operands into a common representation can introduce additional conversion overhead or unnecessary quantization error.
This thesis extends CVFPU, a configurable multi-format floating-point unit, in three directions and integrates the resulting operation groups into the Snitch cluster, a tightly coupled RISC-V compute cluster
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) |
