“Practical NVM/PRAM Usage”: from Ultra-Fast Storage to Deep Learning Accelerator to Checkpoint-Free Persistent Computing
TimeTuesday, June 23rd11:25am - 11:50am
DescriptionData is exploding and generated more in just the last year than in the entire previous history of the industry. As non-volatile memory (NVM) gets closer to processors in modern computing systems, the logical and physical boundaries between them become blurred. While NVM, in particular phase change memory (PRAM), opens the new doors to having persistence and dense memory capacity, it is yet non-trivial to answer a question, "what would be a killer application that make PRAM a real in the near future?" In this talk, we will mainly discuss several applications that we prototyped and innovative concepts to bring NVM/PRAM into a real system. Specifically, this talk advocates different system-level and architecture-level approaches that appropriately leverages the persistence of NVM/PRAM, which includes i) checkpoint-free persistent computing, ii) an ultra-fast PCIe storage card, and iii) a scalable deep learning accelerator. The checkpoint-free computing will show our new processor architecture that integrates RISC-V and NVM/PRAM complex into a single logic fabric, and persistence aware operating system design (OS) while ultra-fast PCIe storage card fully automates all I/O service processes to expose the true latency of NVM/PRAM by taking account for a high energy efficiency. This talk will also demonstrate a deep learning accelerator that takes the advantage of NVM/PRAM's dense memory capacity (rather than persistence). The accelerator can provide a new direction of scalable heterogeneous computing for deep learning acceleration.