NVIDIA GPU and DGX Solutions Guide - Microway

MICROWAY’S NVIDIA® TESLA® V100 GPU SOLUTIONS GUIDE WE SPEAK HPC & AI [email protected] http://microway.com/tesla

WHY MICROWAY? WE SPEAK HPC & AI Many hardware vendors claim GPU expertise, but very few deliver on that promise. Still fewer have been delivering GPU compute since its inception. Microway understands the nuances of GPU hardware - how to architect, build, test, integrate, and deliver it. We’re unique, and we’ll prove why. EXPERT GUIDANCE, INTENSIVE BURN-IN TESTING CUSTOM SOLUTIONS Don’t get stuck with a dud. Every Microway Share the details of your application or code. system receives 72+ hours of burn-in testing prior Microway experts will help you evaluate the best to shipment. This includes specialized stress tests candidate hardware platforms for your application. for NVIDIA GPUs. We find bad components, so Then, they’ll help design a custom configuration you have a superior experience out-of-the box. that’s tuned to your specific needs. COMPLETE GPU SOFTWARE EXPERIENCE: BUILDING GPU INTEGRATION COMPUTING, SINCE THE BEGINNING OF GPU COMPUTING Our team integrates all the drivers, packages, and SDKs that enable you to start working on your We started delivering GPU solutions in 2007—before system or cluster from day 1. For AI deployments, Tesla GPUs even existed. We’ve grown with the we’ll preinstall frameworks—users may even community since, and we have hundreds of satisfied request installation of NVIDIA NGC containers. customers & thousands of GPUs in the field. 2

456 TESLA 1U 4-GPU, OCTOPUTER DGX-2 AI TESLA 4U 4-GPU WITH 8/10 GPUS SUPERCOMPUTER 7 8 9 DGX-1, NVLINK IBM POWER NUMBERSMASHER OCTOPUTER SYSTEM AC922 1U WITH NVLINK 10 11 12 NVIDIA DGX- W H I S P E R S TAT I O N NUMBERSMASHER STATION QUADRO GV100 1U GPU, 1 CPU 13 14 15 NUMBERSMASHER NUMBERSMASHER NAVION 2U GPU 1U, 1 CPU + 4 GPU 8 CPU, 8 GPU WITH 8 GPUS A COMPLETE CATALOG OF GPU PLATFORMS FIND YOUR SOLUTION, OR TALK TO AN EXPERT FOR HELP Microway offers over a dozen GPU platforms to build your cluster, server, or workstation deployment. Few vendors document how platforms are designed, yet how do you identify the best architectures for performance without this information? We’re not afraid of technical details. Overwhelmed? An expert can help steer you to the best solution. 3

BALANCED GPU COMPUTING 2 INTEL XEON CPUS + 4 NVIDIA TESLA GPUS NUMBERSMASHER® 4-GPU 4U NUMBERSMASHER SERVER/TOWER WORKSTATION 1U TESLA 4-GPU SERVER • Flexible form factor • Dense and cost-effective • 4 PCI-E GPUs + 3 additional slots for IB, NVMe • High capacity storage, up to 112TB • 4 PCI-E GPUs + 2 PCI-E x16 slots 4 • InfiniBand and NVMe Storage (Optional)

MAXIMUM GPU CAPACITY OCTOPUTER 4U SERVER WITH A SINGLE PCI-EXPRESS TREE 2 INTEL XEON CPUS + 8 TO 10 NVIDIA TESLA GPUS This Octoputer is available with 8 GPUs + InfiniBand adapters for GPU-Direct RDMA or 10 GPUs for maximum GPU capacity. The single PCI-Express tree ensures low latency and high bandwidth for all GPU peer-to-peer transfers. 5

SCALE UP, WORLD RECORD AI PERFORMANCE NVIDIA DGX-2™ AI APPLIANCE 16 TESLA GPUS, UNIFIED MEMORY WITH NVIDIA NVSWITCH™ 16 fully interconnected Tesla V100 GPUs, 2 TensorPFLOPS and 512GB of unified GPU memory space provide the power to tackle the world’s biggest deep learning and AI challenges. DGX-2 also utilizes NVSwitch and enhanced NVLink technology to ensure seamless data movement—enabling record-breaking performance. 6

SCALE OUT AI PERFORMANCE NVIDIA® TESLA® V100 8-GPU SERVERS WITH NVLINK™ NVIDIA DGX-1™ SUPERCOMPUTER OCTOPUTER™ WITH NVLINK DGX-1 is built and supported by NVIDIA, an ideal With 2 CPUs and 8 NVLink-connected GPUs, solution for deep learning and high performance this Octoputer is ideal for large-memory, data analytics. It includes all the software needed communication-intensive applications. for rapid development and deployment. Optional Mellanox InfiniBand and NVMe storage ensure rapid access to data. 7

WORLD’S SIMPLEST GPU PROGRAMMING POWER SYSTEM AC922 WITH NVIDIA TESLA V100 2 IBM® POWER9® WITH NVLINK CPUS + 4/6 NVIDIA TESLA V100 GPUS The only platform that provides NVLink between CPUs and GPUs—and allows data to flow throughout the system without bottlenecks—while adding full CPU:GPU coherency for the world’s simplest GPU programming. Build your mini- CORAL: these same systems are deployed in the leadership supercomputers at ORNL & LLNL. 8

MAXIMUM NVLINK DENSITY NUMBERSMASHER 1U GPU SERVER WITH NVLINK INTERCONNECT 2 INTEL® XEON® CPUS + 4 NVIDIA TESLA V100 GPUS This system provides the highest GPU density available. With full NVLink connectivity between all GPUs, it enables the highest performance for applications leveraging peer-to-peer GPU communication or GPU-Direct. 9

AI WORKSTATION APPLIANCE NVIDIA DGX-STATION™ INTEGRATED WITH NVIDIA-SUPPORTED SOFTWARE, CONTAINERS DGX-Station with Tesla V100 delivers NVIDIA AI tools, an easy-to-use containerized software platform, and automatic framework updates to deep-learning and AI professionals. Built on fully interconnected Tesla V100 GPUs, DGX-Station provides stunning performance atop of software simplicity. 10

VISUALIZATION, AI, AND COMPUTE WHISPERSTATION™ - QUADRO GV100 4 NVIDIA QUADRO GPUS + 2 XEON CPUS Visualization, deep learning/AI, & computation converge with WhisperStation. Quadro GV100 GPUs provide the actively-cooled equivalent of Tesla V100. WhisperStation is the only customizable professional workstation that provides incredible computational horsepower, unmatched rendering fidelity, and outstanding overall performance in a quiet configuration. Also available with Quadro RTX. 11

COST-EFFECTIVE GPU COMPUTE NUMBERSMASHER 1U 2 GPU, 1 CPU 1 INTEL XEON CPU + 2 NVIDIA TESLA GPUS Deliver the most cost effective density for highly accelerated applications. Pair a cost-effective single CPU socket with 2 Tesla GPUs in a compact 1U footprint. Our most balanced compute node for highly accelerated workloads and a single PCI-E root complex. 12

HIGHLY GPU ACCELERATED NUMBERSMASHER 1U, 4 GPU 1 CPU 1 INTEL XEON CPU + 4 TESLA GPUS For when computation is overwhelmingly offloaded to GPU accelerators, this platform packs 4 GPUs and 1 CPU into a 1U footprint —as much balance towards GPU performance as possible. Deliver the greatest GPU density and utilize GPU-direct RDMA. 13

MAXIMUM MEMORY NUMBERSMASHER 8-CPU + 8-GPU SMP 8 INTEL XEON CPUS + 8 NVIDIA TESLA GPUS Solve huge in-memory computations with up to 12TB system memory! Ideal for many CPU threads plus GPU acceleration, this appliance will provide the highest performance for any data intensive SMP application. 14

HIGHEST GPU:CPU DENSITY IN 1 SERVER NAVION® 2U GPU WITH 8 GPUS 1 AMD EPYC™ CPU + 8 NVIDIA TESLA GPUS The highest GPU:CPU ratio available for overwhelmingly accelerated applications. AMD EPYC’s superior I/O capability enhances accelerated computing. By pairing just a single AMD EPYC CPU with 8 Tesla GPUs, you can allocate the maximum portion of your budget where it delivers most—to the greatest accelerated computing performance boost. 15

EXPERTS IN HIGH PERFORMANCE COMPUTING WWW.MICROWAY.COM Microway designs and manufactures fully-integrated clusters and high performance workstations. For 35 years, we have produced state-of-the-art technical computing solutions for scientists, researchers and engineers. Get in touch with our experts - we take pride in finding the best solution for your HPC needs. 12 Richards Road Plymouth, MA 02360 [email protected] http://microway.com/tesla © Copyright 2018 Microway, Inc. 10-18. All rights reserved. Microway, WhisperStation, NumberSmasher, and Octoputer are trademarks or registered trademarks of Microway, Inc. NVIDIA, DGX-1, DGX Station, Tesla, Pascal, Volta, and NVLink are trademarks or registered trademarks of NVIDIA Corporation. All other trademarks are the property of their respective owners.

HIGHEST GPU:CPU DENSITY IN 1 SERVER NAVION® 2U GPU WITH 8 GPUS 1 AMD EPYC™ CPU + 8 NVIDIA TESLA GPUS The highest GPU:CPU ratio available for overwhelmingly accelerated applications. AMD EPYC’s superior I/O capability enhances accelerated computing. By pairing just a single AMD EPYC CPU with 8 Tesla GPUs, you can allocate the maximum portion of your budget where it delivers most—to the greatest accelerated computing performance boost. 17

WORLD’S SIMPLEST GPU PROGRAMMING POWER SYSTEM AC922 WITH NVIDIA TESLA V100 2 IBM® POWER9® WITH NVLINK CPUS + 4/6 NVIDIA TESLA V100 GPUS The only platform that provides NVLink between CPUs and GPUs—and allows data to flow throughout the system without bottlenecks—while adding full CPU:GPU coherency for the world’s simplest GPU programming. Build your mini- CORAL: these same systems are deployed in the leadership supercomputers at ORNL & LLNL. 18

MAXIMUM NVLINK DENSITY NUMBERSMASHER 1U GPU SERVER WITH NVLINK INTERCONNECT 2 INTEL® XEON® CPUS + 4 NVIDIA TESLA V100 GPUS This system provides the highest GPU density available. With full NVLink connectivity between all GPUs, it enables the highest performance for applications leveraging peer-to-peer GPU communication or GPU-Direct. 19

CONFIGURABLE SCALE OUT NVLINK SYSTEM OCTOPUTER™ WITH NVLINK 2 XEON CPU + 8-GPU SERVER WITH NVLINK™ With 2 CPUs and 8 NVLink-connected GPUs, this Octoputer is ideal for large-memory, communication-intensive applications. Optional Mellanox InfiniBand and NVMe storage ensure rapid access to data. When end-users require a custom-configured system with similar technical architecture to DGX-1/8-GPU NVLink systems, this server is well matched to their requirements. 20

MAX GPUS IN CONFIGURABLE SYSTEM OCTOPUTER - X2 - BUILT ON THE HGX-2 PLATFORM 2 INTEL XEON CPUS + 16 NVIDIA TESLA GPUS Bring the scale of 16 Tesla GPUs, NVSwitch, and 2nd Generation NVLink to your toughest AI challenges. Octoputer-X2 is built upon the NVIDIA HGX-2 platform and offers a single unified 512GB memory space for the largest AI or HPC problems. As a configurable system with similar architecture to DGX-2, it offers flexibility to improve host-memory scale or I/O performance to match your advanced workload. 21

VISUALIZATION, AI, AND COMPUTE WHISPERSTATION™ - QUADRO WITH GV100 OR RTX 6000/8000 4 NVIDIA QUADRO GPUS + 2 XEON CPUS Visualization, deep learning/AI, & computation converge with WhisperStation. Quadro GV100 GPUs provide the actively-cooled equivalent of Tesla V100. WhisperStation is the only customizable professional workstation that provides incredible computational horsepower, unmatched rendering fidelity, and outstanding overall performance in a quiet configuration. Also available with Quadro RTX. 22

HIGH PERFORMANCE AI PROTOTYPING WHISPERSTATION - DEEP LEARNING WITH TITAN RTX ONLY WORKSTATION SUPPORTING 4 TITAN RTX GPUS—WITH CUSTOM LIQUID COOLING Train faster at your desk when prototyping AI models. WhisperStation- Deep Learning integrates up to 4 Liquid Cooled Titan RTX GPUs. This configuration is simply not possible to deploy without Microway’s custom engineered liquid cooling solution. Cool and quiet, this workstation balances performance & cost as you prepare models to deploy on a NVIDIA Tesla, Quadro, or DGX-based server or cluster. 23

EXPERTS IN HIGH PERFORMANCE COMPUTING & AI WWW.MICROWAY.COM Microway designs and manufactures fully-integrated clusters and high performance workstations. For over 35 years, we have produced state-of-the-art technical computing solutions for scientists, researchers and engineers. Get in touch with our experts - we take pride in finding the best solution for your HPC & AI needs. 12 Richards Road Plymouth, MA 02360 [email protected] https://microway.com/tesla https://microway.com/dgx © Copyright 2019 Microway, Inc. 10-19. All rights reserved. Microway, WhisperStation, NumberSmasher, Navion, and Octoputer are trademarks or registered trademarks of Microway, Inc. NVIDIA, DGX-1, DGX-2, DGX Station, Tesla, Volta, and NVLink are trademarks or registered trademarks of NVIDIA Corporation. All other trademarks are the property of their respective owners.

HIGH PERFORMANCE AI PROTOTYPING WHISPERSTATION - DEEP LEARNING WITH TITAN RTX ONLY WORKSTATION SUPPORTING 4 TITAN RTX GPUS—WITH CUSTOM LIQUID COOLING Train faster at your desk when prototyping AI models. WhisperStation- Deep Learning integrates up to 4 Liquid Cooled Titan RTX GPUs. This configuration is simply not possible to deploy without Microway’s custom engineered liquid cooling solution. Cool and quiet, this workstation balances performance & cost as you prepare models to deploy on a NVIDIA Tesla, Quadro, or DGX-based server or cluster. 25

EXPERTS IN HIGH PERFORMANCE COMPUTING & AI WWW.MICROWAY.COM Microway designs and manufactures fully-integrated clusters and high performance workstations. For over 35 years, we have produced state-of-the-art technical computing solutions for scientists, researchers and engineers. Get in touch with our experts - we take pride in finding the best solution for your HPC & AI needs. 12 Richards Road Plymouth, MA 02360 [email protected] https://microway.com/tesla https://microway.com/dgx © Copyright 2020 Microway, Inc. 04-20. All rights reserved. Microway, WhisperStation, NumberSmasher, Navion, and Octoputer are trademarks or registered trademarks of Microway, Inc. NVIDIA, DGX-1, DGX-2, DGX Station, Tesla, Volta, and NVLink are trademarks or registered trademarks of NVIDIA Corporation. All other trademarks are the property of their respective owners.