SC17 Denver, CO

The Internet of Things and HPC: Are They Teaming Up to Work Together?

Authors: Dr. Pete Beckman (Argonne National Laboratory, Northwestern University)

Abstract: IoT devices with edge computing are being deployed everywhere. By almost every physical measure, IoT and HPC seem to be opposites. However, data-driven simulations, machine learning, and in-situ parallel processing are linking them. We will bring together projects for a panel discussion of these questions: a) How will multicore edge devices be programmed? b) Can low-power deep learning accelerators designed for IoT (e.g. Myriad 2) be adapted to HPC? c) Can machine-learning models be shared and connected between IoT and HPC?

Audience: Experts and practitioners in: simulation and modeling, architecture, machine learning, multicore computing, sensing systems, IoT, and mobile computing.

Long Description: HPC and IoT are finding shared interests together. HPC has decades of experience in massively parallel computation, but the IoT ecosystem is rapidly expanding, and the consumer marketplace is driving large-scale investments in hardware and software that is quite similar to HPC. Our goal is to bring together these two communities and look at where exciting technologies and algorithms are shared, and where we can work together on shared research problems. One simple example of the new shared research topics can be seen in the recently announced Movidius (an Intel company) “Neural Compute Stick”, a $79 USB-connected deep learning accelerator based on their Myriad CPU, and used for camera-based gesture recognition in drones. The CPU is extremely low power, with 12 specialized vector VLIW processors. Intel has optimized the deep learning framework Caffe to use the accelerator. The IoT world is rapidly turning to parallel processing for in-situ, edge, computing, and the Myriad 2 is just one of many examples of how hardware and software are overlapping. Another example is the NVIDIA Jetson TX2, designed for embedded and IoT applications. It is 1.5 teraflops Pascal-based system and includes optimized CUDA math libraries for deep neural nets. The overlaps between HPC and IoT are raising many interesting questions, including shared programming models, machine learning frameworks with hardware acceleration, and massive data movement. We will have panel speakers from Intel, ARM, NVIDIA as well as science project currently doing IoT and large-scale instruments.

We held the first, previous BOF at ISC 2017 in Frankfurt, Germany. The room was packed, with standing room only, with more than 100 attendees.

Outcome: We will publish all of the slides, and have started a website for sharing information. We intend to write a very short white paper on the areas of shared research for IoT and HPC.

Conference Presentation: pdf

Birds of a Feather Index