Mastering the three pillars of accelerator design for Deep Learning
Event Type
Special Session (Research Track)
Virtual Programs
Hosted in Virtual Platform
Machine Learning/AI
DescriptionThis talk will present the lessons learnt from designing multiple generations of deep learning accelerators for edge devices to data centers and the three pillars of AI system design: algorithm, hardware, and software. The speaker will describe their industry leading effort in quantization, going from 32 bits down to 8 bits without any loss in training accuracy and down to 4 bits for inference at iso-accuracy. He will also show the micro-architectural design points to exploit the algorithmic advances. Last but not the least, this presentation will cover the compiler and runtime stack, which complements the hardware capabilities.