Agenda S8539 - Pooling and Orchestrating NVIDIA Jetson for AI and Deep Learning on the Edge • Overview ¡of ¡Composable ¡Infrastructure • Fabric ¡and ¡Software ¡Architecture ¡Overview • Jetson ¡Pooling ¡and ¡Scaling ¡Use ¡Case • Summary ¡-‑ Q&A PROPRIETARY & CONFIDENTIAL 2
Agility Always. Orchestrating the Composable Data Center PROPRIETARY & CONFIDENTIAL 3
Liqid Delivers Composable Infrastructure GPU Compute Storage Network Fabric Liqid software and fabric technology enables bare metal provisioning and management of data center resources PROPRIETARY & CONFIDENTIAL – 4/27/2016 4
Impact of Composable PROPRIETARY & CONFIDENTIAL 5
Data Center Roadmap Disaggregated Resources ( Fabric & Software ) PAST PRESENT FUTURE Composable Trillions of Global Traditional Converged & Infrastructure Data Transfers Infrastructure HyperConvrged ( agile, efficient, scalable) PROPRIETARY & CONFIDENTIAL 6
Composable Fabric PCIe is the ONLY common interconnect for ALL data center resources Liqid enables Composable Infrastructure with a PCIe fabric and software that orchestrates and manages bare-metal servers PROPRIETARY & CONFIDENTIAL 7
Software Defined Data Center Server-1 Server-2 Server- N Disaggregated Dynamically Pools of Physical Configure Bare Resources Metal Servers Fabric Storage Pool Compute Pool Graphics Pool Networking Pool PROPRIETARY & CONFIDENTIAL
Jetson Specification Embedded PCIe Interface (Gen2x4) PROPRIETARY & CONFIDENTIAL 9
Jetson Overview Benefits: - Low Power - Small Form Factor Powerful GPU - - Onboard Compute Limitations: Fabric - Limited Storage Limited Networking - - No Standard Deployment Chassis for Scale Out Fabric Based Deployment on Jetson: Storage Pool Networking Pool Enables scale-out in data center environments - - Enables expanded storage - Enables expanded networking PROPRIETARY & CONFIDENTIAL 10
Universal PCIe Expansion Chassis PCIe Fabric Switch 2U 24-Bay Chassis (24x Gen3 x4 Sockets) Networking Storage Off-the-Shelf Off-the-Shelf GPU Networking Storage (Eth, FC, IB, SAS, etc.) (U.2, MLC, 3D, Xpoint, HDD, etc.) Jetson GPU PROPRIETARY & CONFIDENTIAL 11
Universal PCIe Expansion Chassis 2U 24-Bay Chassis (24x Gen3 x4 Sockets) Networking Storage GPU Storage Up ¡to ¡24x ¡U.2 ¡SSD ¡= ¡192 ¡TB Networking Up ¡to ¡48x ¡Networking ¡Ports ¡(Eth, ¡FC, ¡IB, ¡etc..) GPU Up ¡to ¡24x ¡Jetson ¡Modules Management ¡SW Liqid ¡OS ¡Orchestration ¡Software PROPRIETARY & CONFIDENTIAL 12
Dynamic GPU Super Cluster Architecture Overview Machine-1 Machine-2 Machine- N Dynamic Bare Metal Machines Creation Fabric Switch Storage Pool Networking Pool GPU Pool Fabric PROPRIETARY & CONFIDENTIAL 13
Enhanced Jetson Use Cases - Jetson in Edge Servers - Jetson Scale-out in Data Center PROPRIETARY & CONFIDENTIAL 14
Edge Computing Use Case Edge Deployment Requirement: Multi-Server Deployment - - Low Power Small Form Factor - Requires GPU - - Requires Compute Requires 4TB of Storage Per Server - Requires 2x 10Gbe Networking Ports - Option #1) x86 + GPU Server: Option #2) Jetson Server: • 1x Tesla P100 • 1x Jetson TX2 • Intel Xeon 2.5GHz CPU • CPU + GPU Combo • 4TB of SSD • 4TB of SSD • 2x 10Gbe Networking ports • 2x 10Gbe Networking ports PROPRIETARY & CONFIDENTIAL 15
Jetson Server vs. x86 Server Jetson Server x86 + GPU Server GPU 8 W 250 W Compute 0 W 125 W Storage 15 W 15 W Networking 15 W 15 W Total Power 38 W 395 W 90% 90% 85% 85% Lower Lo r Smaller Sm Server Cost $2,750 $13,500 Power Po Footpri Fo rint Form Factor 8x Servers per 2U 1x Server per 2U Watts / Core .029 W .105 W Legacy Servers Jetson Servers Cost / Core $1.17 $3.07 (Fabric Based Server) (2RU Server Chassis) Jetson Enables Lower Power and Smaller Footprint PROPRIETARY & CONFIDENTIAL 16
Jetson in the Data Center Example of Neural Net Model Built/Trained on Tesla and Deployed on Jetson for Inference Jetson deployed on PCIe Fabric can also be used for Training PROPRIETARY & CONFIDENTIAL 17
Inference & Training on Jetson Inference on Jetson in the Field Training on Jetson in the Data Center PROPRIETARY & CONFIDENTIAL 18
Jetson Performance Composed Jetson + NVMe SSD Performance: 1.63 GB/s and 160K IOPS PROPRIETARY & CONFIDENTIAL 19
Summary • Jetson ¡(and ¡Pegasus) ¡are ¡powerful, ¡small ¡form ¡factor • Limited ¡in ¡direct ¡connected ¡storage, ¡networking, ¡and ¡scaling ¡capabilities • Deploying ¡Jetson ¡on ¡a ¡PCIe ¡fabric ¡enables: • Expanded ¡Storage • Expanded ¡Networking • Standard ¡form ¡factor ¡method ¡of ¡deployment • Summary ¡-‑ Q&A PROPRIETARY & CONFIDENTIAL 20
Thank You 21
Recommend
More recommend