Digitala Vetenskapliga Arkivet

Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Design space exploration of in-memory processing architectures
KTH, School of Electrical Engineering and Computer Science (EECS).
2024 (English)Independent thesis Advanced level (degree of Master (Two Years)), 20 credits / 30 HE creditsStudent thesisAlternative title
Designrumsutforskning av bearbetningsarkitekturer i minnet (Swedish)
Abstract [en]

This project focuses on the exploration of analog Processing In Memory (PIM) architectures by leveraging the RAELLA as a baseline architecture model. This thesis extends RAELLA by incorporating real-world physical error models into the design and re-visiting the accuracy estimation. The implemented error models are integrated into the Timeloop/Accelergy simulation framework and used for accuracy evaluations of selected neural network models. We further explore the accuracy behavior under different optimization choices supported in the RAELLA architecture choice. The project addresses two primary research questions. The first question focuses on how to evaluate the accuracy of neural networks while incorporating error models, and the connection between accuracy evaluation and simulation tools, Timeloop/Accelergy. The second question is focused on the trade-offs between accuracy and hardware performance across different neural network configurations, including variations in the depth and width of neural networks. The primary research questions have been addressed with extensive ablation studies performed in a step-by-step fashion to understand the impacts of various optimization choices jointly with memristor device error models. Future studies may focus on exploring more modern memristor devices, implementing and incorporating error models for these devices in the simulator, also addressing the impacts of other non-idealities such as conductance drift on the performance of PIM architectures.

Abstract [sv]

Detta projekt fokuserar på att utforska analoga Processing In Memory (PIM)- arkitekturer genom att utnyttja RAELLA som en baslinjearkitekturmodell. Den här avhandlingen utökar RAELLA genom att införliva fysiska felmodeller i verkligheten i designen och genom att återbesöka noggrannhetsuppskattningen. De implementerade felmodellerna är integrerade i Timeloop/Accelergysimuleringsramverket och används för noggrannhetsutvärderingar av utvalda neurala nätverksmodeller. Vi utforskar ytterligare noggrannhetsbahviour under olika optimeringsval som stöds i RAELLA arkitekturval. Projektet tar upp två primära forskningsfrågor. Den första Frågan fokuserar på hur man utvärderar noggrannheten hos neurala nätverk samtidigt som felmodeller införlivas, och sambandet mellan noggrannhetsutvärdering och simuleringsverktyg, Timeloop/Accelergy. Den andra frågan är fokuserad på avvägningarna mellan noggrannhet och hårdvaruprestanda över olika neurala nätverkskonfigurationer, inklusive variationer i djup och bredd på neurala nätverk. De primära forskningsfrågorna har behandlats med omfattande ablationsstudier utförda i ett steg-för-steg sätt för att förstå effekterna av olika optimeringsval tillsammans med memristor-enhetsfelmodeller. Framtida studier kan fokusera på att utforska toppmoderna memristorenheter, implementera och införliva felmodeller för dessa enheter i simulatorn, och även ta upp effekterna av andra icke-idealiteter såsom konduktansdrift på prestandan hos PIM-arkitekturer.

Place, publisher, year, edition, pages
2024. , p. 49
Series
TRITA-EECS-EX ; 2024:963
Keywords [en]
PIM, crossbar, deep learning, ReRAM
Keywords [sv]
PIM, crossbar, djupinlärning, ReRAM
National Category
Computer and Information Sciences
Identifiers
URN: urn:nbn:se:kth:diva-361380OAI: oai:DiVA.org:kth-361380DiVA, id: diva2:1945220
External cooperation
Ericsson AB
Supervisors
Examiners
Available from: 2025-03-20 Created: 2025-03-18 Last updated: 2025-03-20Bibliographically approved

Open Access in DiVA

fulltext(2088 kB)34 downloads
File information
File name FULLTEXT01.pdfFile size 2088 kBChecksum SHA-512
a08214f676149f9721cc13ab16fdf8c8fe45927035f2ccccc246462843776a032c5edf3e3cf7ac7bd25197bdd5bebadb1f778115373db0b3e74fe5a2ef4cbe15
Type fulltextMimetype application/pdf

By organisation
School of Electrical Engineering and Computer Science (EECS)
Computer and Information Sciences

Search outside of DiVA

GoogleGoogle Scholar
Total: 34 downloads
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

urn-nbn

Altmetric score

urn-nbn
Total: 405 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf