Tutorial EQ10: Advanced Memory and Computing Technologies Using Phase Change Materials

Manuel Le Gallo, IBM Research Europe

Deep Learning, Inference, and Training Using Computational Phase Change Memory

Written by Mohamed Atwa   

Manuel Le Gallo gave a detailed talk on the actualization of deep learning using phase change memory (PCM) hardware. He began by describing the differences between conventional computing, which separates processors and memory and “in-memory” processing, in which computations are done directly in the memory device. He cautioned that only certain types of logic and arithmetic can be done in memory, limited by the device physics in such devices. He then introduced charge and resistance based in-memory computing devices, with DRAM and Flash memory being examples of the former, and ReRAM and PCM being examples of the latter. PCM was given as an example of a prototypical “memristor” for such in-memory processing applications. The use-cases of such memristive “in-memory” processing were presented as being somewhere in between stochastic computing applications such as random number generation and those requiring more exacting solutions, such as scientific computing. Le Gallo then tackled the various nuances surrounding the usage of PCM for deep learning as an ideal use case between stochastic and precise computing. He concluded by introducing a newly-unveiled PCM memory chip simulator that IBM has released to open-source, known as the “IBM Analog AI Hardware Acceleration Kit.”