Architectures and Toolflows for Heterogeneous In-Memory Computing Systems
By minimising data transfer between memory and compute units, in-memory computing has become a viable strategy for increasing the energy efficiency of AI workloads. With an emphasis on accelerator architectures, interconnects, fixed-point near-memory digital processing units, and reconfigurable control infrastructure for deep neural network inference, I will discuss my work on the design of digital and mixed-signal architectures for heterogeneous in-memory computing systems. Using FPGA-based platforms and prototype analogue in-memory computing chips, I will describe the development and assessment of these systems, emphasising architectural trade-offs pertaining to hardware non-idealities, precision, and communication. Programmability and deployment using compiler toolflows to map AI workloads onto heterogeneous hardware platforms will also be covered.
Date and Time
Location
Hosts
Registration
-
Add Event to Calendar
Loading virtual attendance info...
- KTH Electronics and Embedded Systems
- Teknikringen 31
- Stockholm, Stockholms lan
- Sweden 11428
- Building: H-building
- Room Number: 1411
- Click here for Map
Speakers
Dr. Corey Lammie
Architectures and Toolflows for Heterogeneous In-Memory Computing Systems
By minimising data transfer between memory and compute units, in-memory computing has become a viable strategy for increasing the energy efficiency of AI workloads. With an emphasis on accelerator architectures, interconnects, fixed-point near-memory digital processing units, and reconfigurable control infrastructure for deep neural network inference, I will discuss my work on the design of digital and mixed-signal architectures for heterogeneous in-memory computing systems. Using FPGA-based platforms and prototype analogue in-memory computing chips, I will describe the development and assessment of these systems, emphasising architectural trade-offs pertaining to hardware non-idealities, precision, and communication. Programmability and deployment using compiler toolflows to map AI workloads onto heterogeneous hardware platforms will also be covered.