Close

Presentation

Lego: Dynamic Multi-Chip-Module Resource Provision Architecture for Multi-Tenant DNNs
TimeTuesday, July 12th6pm - 7pm PDT
LocationLevel 2 Lobby
Event Type
Networking Reception
Work-in-Progress Poster
DescriptionThe existing DNN accelerators target the acceleration of a single DNN model and limit the throughput for multi-tenant DNN applications. The multi-chip-module (MCM) architecture breaks a monolithic accelerator into multiple small chiplets. However, the high interconnected communication latency on MCMs degrades the performance of multi-tenant DNN workloads significantly. This work proposes Lego MCM architecture that mixes the coarse-grain and fine-grain chiplets to decrease the performance and energy overhead associated with the inter-chip communication. Lego can execute multiple DNNs simultaneously and uses a dynamic resource provision policy to improve the resource utilization and performance of multi-tenant DNN applications on heterogeneous MCMs.