IoT endnodes often couple a small and fast L1 scratchpad memory with higher-capacity but lower bandwidth and speed L2 background memory. The absence of a coherent hardware cache hierarchy saves energy but comes at the cost of labor-intensive explicit memory management, complicating the deployment of algorithms with large data memory footprint, such as Deep Neural Network (DNN) inference. In this work, we present DORY, a lightweight software-cache dedicated to DNN Deployment Oriented to memoRY. DORY leverages static data tiling and DMA-based double buffering to hide the complexity of manual L1-L2 memory traffic management. DORY enables storage of activations and weights in L2 with less than 4% performance overhead with respect to direct execution in L1. We show that a 142 kB DNN achieving 79.9% on CIFAR-10 runs 3.2x faster compared to its execution directly from L2 memory while consuming 1.9x less energy.

Work-in-Progress: DORY: Lightweight Memory Hierarchy Management for Deep NN Inference on IoT Endnodes / Burrello, A; Conti, F; Garofalo, A; Rossi, D; Benini, L. - (2019), pp. 1-2. (Intervento presentato al convegno ESWEEK) [10.1145/3349567.3351726].

Work-in-Progress: DORY: Lightweight Memory Hierarchy Management for Deep NN Inference on IoT Endnodes

Burrello, A;
2019

Abstract

IoT endnodes often couple a small and fast L1 scratchpad memory with higher-capacity but lower bandwidth and speed L2 background memory. The absence of a coherent hardware cache hierarchy saves energy but comes at the cost of labor-intensive explicit memory management, complicating the deployment of algorithms with large data memory footprint, such as Deep Neural Network (DNN) inference. In this work, we present DORY, a lightweight software-cache dedicated to DNN Deployment Oriented to memoRY. DORY leverages static data tiling and DMA-based double buffering to hide the complexity of manual L1-L2 memory traffic management. DORY enables storage of activations and weights in L2 with less than 4% performance overhead with respect to direct execution in L1. We show that a 142 kB DNN achieving 79.9% on CIFAR-10 runs 3.2x faster compared to its execution directly from L2 memory while consuming 1.9x less energy.
2019
9781450369237
File in questo prodotto:
File Dimensione Formato  
ESWEEK_____DORY.pdf

accesso aperto

Tipologia: 2. Post-print / Author's Accepted Manuscript
Licenza: PUBBLICO - Tutti i diritti riservati
Dimensione 1.42 MB
Formato Adobe PDF
1.42 MB Adobe PDF Visualizza/Apri
Pubblicazioni consigliate

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11583/2978562