Search

Your search keyword '"Catron, Garret"' showing total 6 results
Start Over Author "Catron, Garret"
6 results on '"Catron, Garret"'

1. First-Generation Inference Accelerator Deployment at Facebook

Author

Anderson, Michael, Chen, Benny, Chen, Stephen, Deng, Summer, Fix, Jordan, Gschwind, Michael, Kalaiah, Aravind, Kim, Changkyu, Lee, Jaewon, Liang, Jason, Liu, Haixin, Lu, Yinghai, Montgomery, Jack, Moorthy, Arun, Nadathur, Satish, Naghshineh, Sam, Nayak, Avinash, Park, Jongsoo, Petersen, Chris, Schatz, Martin, Sundaram, Narayanan, Tang, Bangsheng, Tang, Peter, Yang, Amy, Yu, Jiecao, Yuen, Hector, Zhang, Ying, Anbudurai, Aravind, Balan, Vandana, Bojja, Harsha, Boyd, Joe, Breitbach, Matthew, Caldato, Claudio, Calvo, Anna, Catron, Garret, Chandwani, Sneh, Christeas, Panos, Cottel, Brad, Coutinho, Brian, Dalli, Arun, Dhanotia, Abhishek, Duncan, Oniel, Dzhabarov, Roman, Elmir, Simon, Fu, Chunli, Fu, Wenyin, Fulthorp, Michael, Gangidi, Adi, Gibson, Nick, Gordon, Sean, Hernandez, Beatriz Padilla, Ho, Daniel, Huang, Yu-Cheng, Johansson, Olof, Juluri, Shishir, Kanaujia, Shobhit, Kesarkar, Manali, Killinger, Jonathan, Kim, Ben, Kulkarni, Rohan, Lele, Meghan, Li, Huayu, Li, Huamin, Li, Yueming, Liu, Cynthia, Liu, Jerry, Maher, Bert, Mallipedi, Chandra, Mangla, Seema, Matam, Kiran Kumar, Mehta, Jubin, Mehta, Shobhit, Mitchell, Christopher, Muthiah, Bharath, Nagarkatte, Nitin, Narasimha, Ashwin, Nguyen, Bernard, Ortiz, Thiara, Padmanabha, Soumya, Pan, Deng, Poojary, Ashwin, Ye, Qi, Raginel, Olivier, Rajagopal, Dwarak, Rice, Tristan, Ross, Craig, Rotem, Nadav, Russ, Scott, Shah, Kushal, Shan, Baohua, Shen, Hao, Shetty, Pavan, Skandakumaran, Krish, Srinivasan, Kutta, Sumbaly, Roshan, Tauberg, Michael, Tzur, Mor, Verma, Sidharth, Wang, Hao, Wang, Man, Wei, Ben, Xia, Alex, Xu, Chenyu, Yang, Martin, Zhang, Kai, Zhang, Ruoxi, Zhao, Ming, Zhao, Whitney, Zhu, Rui, Mathews, Ajit, Qiao, Lin, Smelyanskiy, Misha, Jia, Bill, and Rao, Vijay

Subjects
Computer Science - Hardware Architecture
Abstract

In this paper, we provide a deep dive into the deployment of inference accelerators at Facebook. Many of our ML workloads have unique characteristics, such as sparse memory accesses, large model sizes, as well as high compute, memory and network bandwidth requirements. We co-designed a high-performance, energy-efficient inference accelerator platform based on these requirements. We describe the inference accelerator platform ecosystem we developed and deployed at Facebook: both hardware, through Open Compute Platform (OCP), and software framework and tooling, through Pytorch/Caffe2/Glow. A characteristic of this ecosystem from the start is its openness to enable a variety of AI accelerators from different vendors. This platform, with six low-power accelerator cards alongside a single-socket host CPU, allows us to serve models of high complexity that cannot be easily or efficiently run on CPUs. We describe various performance optimizations, at both platform and accelerator level, which enables this platform to serve production traffic at Facebook. We also share deployment challenges, lessons learned during performance optimization, as well as provide guidance for future inference hardware co-design.

Published
2021

2. Glow: Graph Lowering Compiler Techniques for Neural Networks

Author

Rotem, Nadav, Fix, Jordan, Abdulrasool, Saleem, Catron, Garret, Deng, Summer, Dzhabarov, Roman, Gibson, Nick, Hegeman, James, Lele, Meghan, Levenstein, Roman, Montgomery, Jack, Maher, Bert, Nadathur, Satish, Olesen, Jakob, Park, Jongsoo, Rakhov, Artem, Smelyanskiy, Misha, and Wang, Man

Subjects
Computer Science - Programming Languages
Abstract

This paper presents the design of Glow, a machine learning compiler for heterogeneous hardware. It is a pragmatic approach to compilation that enables the generation of highly optimized code for multiple targets. Glow lowers the traditional neural network dataflow graph into a two-phase strongly-typed intermediate representation. The high-level intermediate representation allows the optimizer to perform domain-specific optimizations. The lower-level instruction-based address-only intermediate representation allows the compiler to perform memory-related optimizations, such as instruction scheduling, static memory allocation and copy elimination. At the lowest level, the optimizer performs machine-specific code generation to take advantage of specialized hardware features. Glow features a lowering phase which enables the compiler to support a high number of input operators as well as a large number of hardware targets by eliminating the need to implement all operators on all targets. The lowering phase is designed to reduce the input space and allow new hardware backends to focus on a small number of linear algebra primitives.

Published
2018

4. Tip-over Prevention Through Heuristic Reactive Behaviors for Unmanned Ground Vehicles

Author

SPACE AND NAVAL WARFARE SYSTEMS CENTER PACIFIC SAN DIEGO CA, Talke, Kurt, Kelley, Leah, Longhini, Patrick, Catron, Garret, SPACE AND NAVAL WARFARE SYSTEMS CENTER PACIFIC SAN DIEGO CA, Talke, Kurt, Kelley, Leah, Longhini, Patrick, and Catron, Garret

Abstract

Skid-steer teleoperated robots are commonly used by military and civilian crews to perform high-risk, dangerous and critical tasks such as bomb disposal. Their missions are often performed in unstructured environments with irregular terrain covered with a variety of media, such as sand, brush, mud, rocks and debris. During such missions, it is often impractical if not impossible to send another robot or a human operator to right a toppled robot. As a consequence, a robot tip-over event usually results in mission failure. Also such robots are often equipped with heavy payloads that raise their centers of mass and hence increase their instability. Should the robot be equipped with a manipulator arm or flippers, it may have a way to self-right. The majority of manipulator arms are not designed for and are likely to be damaged during self-righting procedures, however. Those robots not equipped with manipulator arms or flippers have no self-righting capabilities. Due to the on-board camera frame of reference, the video feed may cause the robot to appear to be on at level ground when it actually may be on a slope nearing tip-over. Finally, robot operators are often so focused on the mission at hand they are oblivious to their surroundings. Our research seeks to remove tip-over monitoring from the already large list of tasks an operator must perform. An autonomous tip-over prevention behavior for a mobile robot with a static payload has been developed implemented and experimentally validated on two different teleoperated robotic platforms. Suitable for use with both teleoperated and autonomous robots, the prevention behavior uses the force-angle stability measure previously experimentally validated, to predict the likelihood of robot tip-over and trigger prevention behaviors., Presented at SPIE Defense, Security, and Sensing, DSS 2014, in Baltimore, MD on 6-8 May 2014 and published in SPIE Proceedings v9084: Unmanned Systems Technology XVI.

Published
2014

Catalog

Books, media, physical & digital resources
See catalog results