A CoSimulation Framework for Assessment of Power Knobs in Deep-Learning Accelerators
MetadataShow full item record
This thesis provides a tool able to meet the demands of a collaborative work between machine-learning experts and digital designers, with a particular attention to hardware accelerators implementing a spatial architecture. The latter consists of a large number of processing elements, interconnected with a network-on-chip allowing the sharing of operands and to carry on computations spatially. In particular, the contribution of this work has consists in the development of a co-simulation framework able to: • Evaluate the effective energy efficiency of realistic workload of spatial accelerators avoiding the simulation of the entire accelerator micro-architecture. • Explore the design space to evaluate pros and cons of the designated HW architecture and power-management strategy. • Enable an early efficiency testing of energy-aware neural network model, without waiting for the complete design of the accelerator, thanks to an accurate estimation of the energy profile of the real hardware platform. The tool has been designed in order to be easily interfaced with common frameworks for machine learning and with the industrial ASIC design flow. The general philosophy behind the co-simulation framework is to have a behavioral neural network inferential engine that communicates with a gate-level simulator: the inferential engine can provide stimuli to the circuit, collect responses, status signals and modify the configuration of power knobs. The aim is to simulate the system also from a non-functional perspective, thus the need for a gate level simulator, but with only the minimum hardware required to verify the impact of a specific power management strategy on the network accuracy. In particular, the effect of power knobs on the system is emulated through a library of SDF files, one for each working condition, which can be loaded by the gate-level simulator when a power-context switch is performed.
Computer Architectures for Deep Learning
Energy Efficient Deep Learning