Theme 1: Heterogeneous Computing Platforms

We envision a planet-scale distributed computing infrastructure with a myriad of heterogeneous accelerators. Accelerators will rapidly evolve with applications and, in addition, at any point in time, co-exist with earlier or later generations. Hence, we propose a new methodology to easily generate, deploy, and reconfigure Evolvable accelerators. Groups of accelerators will be organized into Ensembles distributed across one or multiple datacenters.  Applications will dynamically pick (and reconfigure) the desired set of accelerators from the ensemble with minor overhead. Advanced runtime and compilation methods will reconfigure multi tenant accelerator ensembles, and map  and schedule applications to them. Finally, revamped general-purpose cores will differentiate to increase performance and energy efficiency.

Inside Theme 1-2
Design-space exploration of reconfigurable ASIC accelerators (Courtesy of Zhiru Zhang).

Papers and Presentations:

SPADE: A Flexible and Scalable Accelerator for SpMM and SDDMM
Gerasimos Gerogiannis, Serif Yesil, Damitha Lenadora, Dingyuan Cao, Charith Mendis, Josep Torrellas
International Symposium on Computer Architecture (ISCA), June 2023
MXFaaS: Resource Sharing in Serverless Environments for Parallelism and Efficiency
Jovan Stojkovic, Tianyin Xu, Hubertus Franke, Josep Torrellas
International Symposium on Computer Architecture (ISCA), June 2023
μManycore: A Cloud-Native CPU for Tail at Scale
Jovan Stojkovic, Chunao Liu, Muhammad Shahbaz, Josep Torrellas
 International Symposium on Computer Architecture (ISCA), June 2023