Grant: Research Tools and Instruments

This equipment grant supports purchase of an AI supercomputer containing tightly-coupled graphics processing units (GPUs), which will be used for expediting research funded by two large NSERC Strategic Partnership Grants. The first is NSERC Strategic Networks grant NETGP485577-15 entitled, Computing Hardware for Emerging Intelligent Sensory Applications (COHESA). The second is NSERC Strategic Project grant STPGP 506681-17 entitled, Error resilient machine learning systems. Both projects are related to improving hardware for machine learning systems as briefly outlined below. NSERC COHESA is motivated by the need and the opportunity to advance the state-of-the-art in ISAs, which are helping computing devices to ""think"", ""see"", ""hear"", ""read"", ""write"" and ""interact"" with the physical world in ways that we typically associate only with humans and high intelligence. However, ISA is hampered by the limitations of existing computing system. The focus of COHESA is developing novel techniques at the hardware, system software and algorithm layers to improve the effectiveness of machine learning systems by achieving orders of magnitude reductions in hardware costs in spite of slowing semiconductor technology advances. Using the equipment purchased with this grant we will characterize the redundancies in emerging training and inference algorithms that will provide the insights needed to optimize future hardware and software for deep neural networks. Our work on error resilient machine learning systems is motivated by the fact that an important source of faults is due to errors in computer chips. These errors, also known as soft errors, occur due to cosmic rays and other particles hitting the electronic devices on the chips, causing them to malfunction. Soft errors are increasing in frequency becausechip manufacturers move towards smaller devices so that they can pack more of them on the chip, thereby making the chips faster. Soft errors can compromise the safety of machine learning applications such as robotics, and need to be mitigated. Using the equipment supported by this grant we will perform detailed characterizations of soft errors in the latest machine learning enabled GPUs using the radiation testing facilities at TRIUMF located on UBC campus. We will use the knowledge from these experiments to propose machine learning hardware and software that is more resilient to soft errors.