Latest HPC News from NVIDIA
1 like826 views
The document discusses NVIDIA's role in powering the world's fastest supercomputers. It notes that the Summit supercomputer at Oak Ridge National Laboratory is now the fastest system, powered by 27,648 Volta Tensor Core GPUs to achieve over 122 petaflops. NVIDIA GPUs also power 17 of the world's 20 most energy efficient supercomputers, including Europe's fastest Piz Daint and Japan's fastest Fugaku supercomputer. Over 550 applications are now accelerated using NVIDIA GPUs.
Latest HPC News from NVIDIA
- 1. Duncan Poole, NVIDIA ISC 2018
- 2. 2 NVIDIA POWERS WORLD'S FASTEST SUPERCOMPUTER 27,648 Volta Tensor Core GPUs Summit Becomes First System To Scale The 100 Petaflops Milestone 122 PF 3 EF HPC AI
- 3. 3 NVIDIA POWERS FASTEST SUPERCOMPUTERS IN US, EUROPE, JAPAN, INDUSTRY 17 of World’s 20 Most Energy-efficient Supercomputers Piz Daint Europe’s Fastest 5,320 GPUs| 20 PF ORNL Summit World’s Fastest 27,648 GPUs| 122 PF ABCI Japan’s Fastest 4,352 GPUs| 20 PF ENI HPC4 Fastest Industrial 3,200 GPUs| 12 PF LLNL Sierra US 2nd Fastest 17,280 GPUs| 72 PF
- 4. 4 ALL TOP 15 APPLICATIONS ACCELERATED 550+ Applications Accelerated 8X CUDA DOWNLOADS 2018 8M 1M 2012 DEFINING THE NEXT GIANT WAVE IN HPC OAK RIDGE SUMMIT World’s fastest supercomputer 120+ Petaflop HPC; 3+ Exaflop of AI ABCI Supercomputer (AIST) Japan’s fastest AI supercomputer Piz Daint Europe’s fastest supercomputer MOST ADOPTED PLATFORM FOR ACCELERATING HPC 259 319 400 470 554 2014 2015 2016 2017 2018 #of GPU-Acc elerat ed Apps
- 5. 5 NVIDIA SDK & LIBRARIES INDUSTRY FRAMEWORKS & APPLICATIONS CUSTOMER USECASES SUPERCOMPUTING +550 Applications CUDA NCCLcuDNN TensorRTcuBLAS DeepStreamcuSPARSEcuFFT Amber NAMDLAMMPS CHROMA ENTERPRISE APPLICATIONSCONSUMER INTERNET ManufacturingHealthcare EngineeringSpeech Translate Recommender Molecular Simulations Weather Forecasting Seismic Mapping cuRAND NVIDIA TESLA PLATFORM World’s Leading Data Center Platform for Accelerating HPC and AI TESLA GPUs & SYSTEMS SYSTEM OEM CLOUDTESLA GPU NVIDIA HGXNVIDIA DGX FAMILY
- 6. 6 END-TO-END PRODUCT FAMILY HPC/TRAINING INFERENCE EMBEDDED Jetson TX1 DATA CENTER Tesla P4 AUTOMOTIVE Drive PX2 Tesla P100Tesla V100Titan V DATA CENTERDESKTOP FULLY INTERGRATED DL SUPERCOMPUTER Tesla V100TITAN V DESKTOP WORKSTATION DATA CENTER Tesla V100 DGX StationTITAN Quadro DGX-1 DGX-2 V100 PCIE FULLY INTEGRATED AI SYSTEMS
- 7. 7 GPU-Accelerated Server Platform Dell EM C Fujitsu HPE IBM Lenovo Superm icro SCX-E4 › 4x V100 NVLINK • Pow erEdge C4140 • Prim ergy CX400 M4 • Pow er System s AC922* • SYS-1028GQ-TVRT SCX-E3 › 8x V100 PCIE • Apollo 6500 (XL270d Gen10) SCX-E2 › 4x V100 PCIE • Pow erEdge C4140 • Pow erEdge T640 • Pow erEdge R940xa • Prim ergy CX400 M4 • SD530/D2 • SYS-1029GQ SCX-E1 › 2x V100 PCIE • Pow erEdge R840 • Pow erEdge R740xd* • Pow erEdge R740* • Prim ergy RX2540 M4 • SD650 HGX-T1 › 8x V100 NVLINK • Apollo 6500 (XL270d Gen10) • SYS-4029GP-TVRT V100 32GB SERVERS AVAILABLE FROM OEMS Server Catalog *reduced GPU configuration
- 8. 8 TESLA V100 Form Factor Performance 7.8TF DP, 15.7TF SP, 125TF FP16 7TF DP, 14TF SP, 112TF FP16 Memory Size 16GB/32GB HBM2 Memory Bandwidth 900GB/s GPU Peer to Peer NVLink (up to 300 GB/s) + PCIe Gen3 (up to 32 GB/s) PCIe Gen3 (up to 32 GB/s) Power 300W 250W Available From All Major OEMs S X M 2 3 2 G B P / N = 9 0 0 - 2 G 5 0 3 - 0 0 1 0 - 0 0 0 , P C IE 3 2 G B P / N = 9 0 0 - 2 G 5 0 0 - 0 0 1 0 - 0 0 0 , S X M 2 1 6 G B P / N = 9 0 0 - 2 G 5 0 3 - 0 0 0 0 - 0 0 0 , P C IE 1 6 G B P / N = 9 0 0 - 2 G 5 0 0 - 0 0 0 0 - 0 0 0 SXM2 PCIe
- 9. 9 NVIDIA DGX-2 AND HGX-2
- 10. 10 NVSWITCH WORLD’S HIGHEST BANDWIDTH ON-NODE SWITCH 7.2 Terabits/sec or 900 GB/sec 18 NVLINK ports | 50GB/s per port bi-directional Fully-connected crossbar 2 billion transistors | 47.5mm x 47.5mm package
- 11. 11 NVSWITCH ENABLES THE WORLD’S LARGEST GPU 16 Tesla V100 32GB Connected by New NVSwitch 2 petaFLOPS of DL Compute Unified 512GB HBM2 GPU Memory Space 300GB/sec Every GPU-to-GPU 2.4TB/sec of Total Cross-section Bandwidth
- 12. 12 THE LARGEST, FASTEST SHARED MEMORY SUPERNODE FOR THE MOST DIFFICULT HPC CHALLENGES • 125 TFLOPS DP, 250 TF SP • 512 GB shared memory • 14.4 TB/s aggregate HBM BW • 2.4 TB/s bisection BW • 8x EDR network • 30 TB SSD INTRODUCING NVIDIA DGX-2 THE WORLD’S MOST POWERFUL HPC SUPERNODE
- 13. 13 INSIDE DGX-2: “WORLD’S LARGEST GPU” 1 2 3 5 4 6 Two Intel Xeon Platinum CPUs 7 1.5 TB System Memory 13 30 TB NVME SSDs Internal Storage NVIDIA Tesla V100 32GB Two GPU Boards 8 V100 32GB GPUs per board 6 NVSwitches per board 512GB Total HBM2 Memory interconnected by Plane Card Twelve NVSwitches 2.4 TB/sec bi-section bandwidth Eight EDR Infiniband/100 GigE 1600 Gb/sec Total Bi-directional Bandwidth PCIe Switch Complex 8
- 14. 14 OVER 2X HIGHER PERFORMANCE WITH NVSWITCH Two DGX-1 Compared to DGX-2 2 H G X - 1 V s e r v e r s h a v e d u a l s o c k e t X e o n E 5 2 6 9 8 v 4 P r o c e s s o r . 8 x V 1 0 0 G P U s . S e r v e r s c o n n e c t e d v ia 4 X 1 0 0 G b IB p o r t s ( r u n o n D G X - 1 ) | H G X - 2 s e r v e r h a s d u a l- s o c k e t X e o n P la t in u m 8 1 6 8 P r o c e s s o r . 1 6 V 1 0 0 G P U s ( r u n o n D G X - 2 ) Physics (MILC benchmark) 4D Grid Weather (ECMWF benchmark) All-to-all Recommender (Sparse Embedding) Reduce & Broadcast Language Model (Transformer with MoE) All-to-all DGX-2 with NVSwitchTwo DGX-1 (Volta) 2X FASTER 2.4X FASTER 2X FASTER 2.7X FASTER AI TrainingHPC
- 15. 15 TESLA HGX-2 FUSING HPC AND AI INTO ONE UNIFIED COMPUTING ARCHITECTURE Multi-precision Computing 2 PFLOPS AI | 250 TFLOPS FP32 | 125 TFLOPS FP64 16 Tesla V100 GPUs | 0.5TB Memory | 2.4 TB/s Building Block for Partner Systems & DGX-2
- 16. 16 NVIDIA SOFTWARE PLATFORM UPDATES
- 17. 17 NVIDIA GPU CLOUD (NGC) Simple Access to GPU-Accelerated Software Cloud Servers Workstations Deploy Applications In Minutes, Not Days Discover 35 Optimized Containers Run Anywhere with Maximum Performance GPU-Powered Accelerate Time to Market
- 18. 18 CONTAINERS SIMPLIFY APPLICATION DEPLOYMENTS DRIVERS + OPERATING SYSTEM CONTAINER RUNTIME NAMD 2.12 CUDA libraries VMD CUDA libraries GROMACS CUDA libraries NAMD 2.13 CUDA libraries Environment modules simplified/eliminated Performance equivalent to bare metal Deploy applications in minutes Higher productivity for sys admins & users SHARED CLUSTER Portable on various systems Reproducible results
- 19. 19 NGC CONTAINER REGISTRY 10 Containers at Launch, 35 Containers Today bigdft candle chroma gamess gromacs lammps lattice-microbes MILC namd pgi picongpu relion vmd caffe caffe2 cntk cuda digits inferenceserver mxnet pytorch tensorflow tensorrt theano torch index paraview-holodeck paraview-index paraview-optix chainer h20ai-driverless kinetica mapd paddlepaddle Deep Learning HPC HPC Visualization PartnersNVIDIA/K8s Kubernetes on NVIDIA GPUs *NewContainers since SC 17
- 20. 20 CUDA TOOLKIT 9.2 Optimized for Volta: • Tensor Cores • Second-Generation NVLink • HBM2 Stacked Memory UNLEASHES POWER OF VOLTA COOPERATIVE THREAD GROUPS Flexible Thread Groups Efficient Parallel Algorithms • Synchronize Across Thread Blocks in a Single GPU or Multi-GPUs • RNN and CNN Optimizations (cuBLAS) • >20x Faster Image Processing (NPP) • Speed up FFT of prime size matrices (cuFFT) FASTER LIBRARIES DEVELOPER TOOLS & PLATFORM UPDATES • CUTLASS 1.0 accelerate custom linear algebra algorithms • 2x faster CUDA kernel launch • New OS & Compiler Support • Unified Memory Profiling • NVLink Visualization
- 21. 21 CUDA 9.2 PLATFORM SUPPORT New OS and Host Compilers PLATFORM OS VERSION COMPILERS Windows Windows Server 2016 2012 R2 Microsoft Visual Studio 2017 (15.6) Linux 16.04.4 LTS 17.10 non GCC 7.x PGI 18.x Clang 5.0.x ICC 17 XLC 13.1.6 (POWER) 7.5 7.5 POWER LE SLES 12 SP3 27 Leap 42.3 Mac macOS 10.13.4 Xcode 9.2
- 22. 22 silica IFPEN, RMM-DIIS on P100 OPENACC GROWING MOMENTUM Wide Adoption Across Key HPC Codes ANSYS Fluent Gaussian VASP LSDalton MPAS GAMERA GTC XGC ACME FLASH COSMO Numeca Over 100 Apps* Using OpenACC Prof. Georg Kresse Computational Materials Physics University of Vienna For VASP, OpenACC is the way forward for GPU acceleration. Performance is similar to CUDA, and OpenACC dramatically decreases GPU development and maintenance efforts. We’re excited to collaborate with NVIDIA and PGI as an early adopter of Unified Memory. VASP Top Quantum Chemistry and Material Science Code * Applications in production and development
- 23. 23 DCGM Active Health Monitoring • Run-time health checks • Prologue check: Quick health check of the GPU • Epilogue check: online GPU diagnostic tests to determine root cause issues NVIDIA Data Center GPU Manager developer.nvidia.com/cuda-toolkit Diagnostics & System Validation • GPU Compute Performance • Interconnect BW & Latency • Power & Thermals Policy Framework • Assists in Recovery Action Automation • Group Control over Power & Clock Policy • Dynamic Page Retirement Policy