Close

Presentation

Morphable CIM: Improving Operation Intensity and Depthwise Capability for SRAM-CIM Architecture
DescriptionSRAM-CIM achieves better weight operation intensity than Von Neumann architecture. However, sticking on weight-stationary dataflows is suboptimal since some layers of CNNs suffer from low activation operation intensity and underutilization.

Morphable CIM contains the following key contributions:
1) A dataflow-morphable architecture that switches between weight-stationary and input-stationary dataflow.
2) An input-stationary systolic dataflow and a word-wise mapping.
3) An window-unrolled input mapping and a depthwise-capable CIM macro to improve utilization for depthwise layers.

The experimental results show that our proposed SRAM-CIM architecture outperforms the baselines by up to 16.7x performance speedup, 3.6x energy efficiency, and 94.8% memory traffic reduction.
Event Type
Research Manuscript
TimeWednesday, July 12th4:10pm - 4:25pm PDT
Location3004, 3rd Floor
Topics
Design
Keywords
In-memory and Near-memory Computing Circuits