Research Experience
-
Hardware/Software co-design of quantization for Nerual Network
Work in progress
Advisor: Prof. Xinfei Guo
Duration: Oct 2021 - Present- Hardware-Software Co-design
Support the post-training MPQ with tightly integrated vector-matrix multiply units in a CPU pipeline through micro-architecture and ISA co-design. - A Novel MPQ Searching Algorithm
Propose new metric $w$ and use the greedy algorithm to sample the layer-wise sensitivity with respect to both accuracy and proxy of hardware cost. Propose a metric denoted as $\mathcal{P}$ to measure the influence of QAT on accuracy over PTQ. - Impact
Proposed search algorithm achieves $3%\sim11%$ higher inference accuracy with similar SOTA MPQ strategies.
- Hardware-Software Co-design
-
High-Quality Macro Placement with Enhanced Dataflow-Aware Guidance
Work in progress
Advisor: Prof. Xinfei Guo
Duration: Aug 2022 - Present- Based on Open Source EDA platform - OpenROAD
Research of macro placement based on open source EDA software OpenROAD. - Dataflow-Aware Connection Between Macro and Standard Cell
Use virtual weighted connections to establish data flow relationships between macro and macro and between macro and cell. - Impact
Compared with other macro placer in OpenROAD, HPWL has better optimization effect.
- Based on Open Source EDA platform - OpenROAD
