2021 58th ACM/IEEE Design Automation Conference (DAC) 2021
DOI: 10.1109/dac18074.2021.9586312
|View full text |Cite
|
Sign up to set email alerts
|
Help me understand this report

Search citation statements

Order By: Relevance

Paper Sections

Select...
2
1
1
1

Citation Types

0
11
0

Year Published

2021
2021
2023
2023

Publication Types

Select...
5
1

Relationship

0
6

Authors

Journals

citations
Cited by 19 publications
(11 citation statements)
references
References 12 publications
0
11
0
Order By: Relevance
“…The first line of work has focused on designing reconfigurable architectures that allow the dynamic composition of engines to build larger processing blocks (Figure 2b). This technique can be applied either by combining coarse systolic arrays with a uniform shape [5] or at a finer-grained manner using varying-sized systolic arrays [6]. In both of these cases, resource partitioning decisions are made at run time based on the DNNs that are co-located at any time instant.…”
Section: Spatial Co-locationmentioning
confidence: 99%
See 3 more Smart Citations
“…The first line of work has focused on designing reconfigurable architectures that allow the dynamic composition of engines to build larger processing blocks (Figure 2b). This technique can be applied either by combining coarse systolic arrays with a uniform shape [5] or at a finer-grained manner using varying-sized systolic arrays [6]. In both of these cases, resource partitioning decisions are made at run time based on the DNNs that are co-located at any time instant.…”
Section: Spatial Co-locationmentioning
confidence: 99%
“…Despite its merits, the arrays' uniform size restricts the customization potential and can lead to underutilization. To counteract this, FGSpMt-NPU [6] proposed the dataflow-mirroring technique. Instead of composing fixed-sized arrays into larger structures, this scheme combines individual rows and columns of the arrays to build arbitrarily sized blocks.…”
Section: Enhancing Existing Architecturesmentioning
confidence: 99%
See 2 more Smart Citations
“…In this manner, the different sublayers complementarily utilise the available computational and bandwidth resources, leading to high performance and efficient sharing of the proposed accelerator. Similarly, [81] and [85] also target single DNN engines and present dataflow mirroring and a preemption module, respectively, two hardware-level enhancements that aim to optimise the concurrent execution of multiple co-located DNNs on the underlying engine.…”
Section: A Multi-tenant Ai Systemsmentioning
confidence: 99%