Tech Talk #18 – Neural Network Accelerators: From Physics to System with a Cool Approximation

Την Παρασκευή 4/2/2021 και ώρα 10:30 EEST (Greek Time) θα έχουμε την παρουσίαση για τα Fridays Tech Talks από τoν μεταδιδάκτορα Γεώργιο Ζερβάκη με τίτλο:

“Neural Network Accelerators: From Physics to System with a Cool Approximation”

Η παρουσίαση θα γίνει μέσω webex στο παρακάτω link Webex.

Ακολουθεί μία σύντομη περίληψη της παρουσίασης:

“Modern systems, from the ultra-low power to the HPC domain, rely more and more on deep neural networks (DNNs) to deliver sophisticated services. To satisfy tighter and tighter latency requirements, ASIC inference accelerators are becoming mainstream. Inference accelerators perform thousands of MAC operations per cycle, packing immense computational power in a very confined area, leading thus to elevated power densities and unsustainable on-chip temperature. To further boost the performance of DNN accelerators while satisfying strict temperature and energy constraints, radical departures to conventional approaches are mandatory. In this talk, we present a holistic evaluation, from the device level all the way up to the system level where inference accuracy is affected, of DNN accelerators. We evaluate the impact that emerging technology (NCFET), emerging cooling (Superlattice Thermoelectric Cooling), and emerging computing (Approximate Computing) have on future DNN accelerators in terms of accuracy, efficiency, temperature, and/or performance.”