Session
When They Go High - We Go Low (in Computation)
Event TypeResearch Manuscript
Presented In-Person
Approximate Computing for AI/ML
Design
TimeTuesday, December 7th3:30pm - 5:00pm PST
Location3014
DescriptionLow computational complexity improves hardware efficiency. To that end, this session reports cutting-edge developments in the domain of Approximate Computing for AI Hardware-Software codesign. Along with novel algorithmic techniques addressing pruning and reduced precision, the session dives deeper into cross-layer optimization and approximate arithmetic for different types of deep neural network accelerators ranging from embedded FPGA systems to in-memory computing.
Presentations
| 3:30pm - 3:50pm PST | Softermax: Hardware/Software Co-Design of an Efficient Softmax for Transformers | |
| 3:50pm - 4:10pm PST | Control Variate Approximation for DNN Accelerators | |
| 4:10pm - 4:30pm PST | COSAIM: Counter-based Stochastic-behaving Approximate Integer Multiplier for Deep Neural Networks |