Paper title:
ALRESCHA: A Lightweight Reconfigurable Sparse-Computation Accelerator
Publication:
HPCA’20
Problem to solve:
The parallelism and data dependency problem in sparse computation workloads.
Major contribution:
-
Proposed a generic sparse accelerator for both scientific calculations and graph calculations regardless of whether they have data-dependent patterns.
-
To support the accelerator mentioned above, this paper proposed a lightweight reconfigurability method, which reconfigures the part of the accelerator runtime.
-
To support the whole method, this paper proposed a storage format to stream data and facilitate the computations with data-dependency.
Lessons learnt:
-
This paper actually composes the methods in two kinds of accelerators: GEMV computation accelerator and graph computation accelerator. However, the implementation of fast runtime switch between configurations in RCU is still valuable.
-
The GEMV accelerator part proposed by this paper is not much innovative, while the most innovative part is the storage part. This kind of format helps the graph computations most.