Emergent Mind

Abstract

To meet the ever-increasing computation demand from emerging workloads, a scalable design paradigm combines multiple Deep Neural Network (DNN) accelerators to build a large multi-accelerator system. They are mainly proposed for data centers, where workload varies across vision, language, recommendation, etc. Existing works independently explore their hardware configuration and mapping strategies due to the extremely large cross-coupled design space. However, hardware and mapping are interdependent and, if not explored together, may lead to sub-optimal performance when workload changes. Moreover, even though a data center accelerator has multiple objectives, almost all the existing works prefer aggregating them into one (mono-objective). But aggregation does not help if the objectives are conflicting, as improving one will worsen the other. This work proposes MOHaM, a multi-objective hardware-mapping co-optimisation framework for multi-tenant DNN accelerators. Specifically, given an application model and a library of heterogeneous, parameterised and reconfigurable sub-accelerator templates, MOHaM returns a Pareto-optimal set of multi-accelerator systems with an optimal schedule for each one of them to minimise the overall system latency, energy and area. MOHaM is evaluated for diverse workload scenarios with state-of-the-art sub-accelerators. The Pareto-optimal set of competitive design choices enables selecting the best one as per the requirement.

We're not able to analyze this paper right now due to high demand.

Please check back later (sorry!).

Generate a summary of this paper on our Pro plan:

We ran into a problem analyzing this paper.

Newsletter

Get summaries of trending comp sci papers delivered straight to your inbox:

Unsubscribe anytime.