Intelligent Interconnect Architecture to Enable Next Generation HPC - Linaro HPC Workshop 2018
2 likes2,382 views
Mellanox Technologies highlights its role in accelerating high-performance computing (HPC) and artificial intelligence (AI) systems with advanced interconnect solutions, including the world's fastest supercomputers using Infiniband technology. It emphasizes the importance of data-centric architecture and in-network computing to overcome latency bottlenecks, achieving significant performance improvements. The document discusses various hardware innovations, including the BlueField system on a chip, designed to optimize networking and processing capabilities in data centers.
Intelligent Interconnect Architecture to Enable Next Generation HPC - Linaro HPC Workshop 2018
- 1. 1© 2018 Mellanox Technologies | Confidential Paving the Road to Exascale July 2018 Interconnect Your Future
- 2. 2© 2018 Mellanox Technologies | Confidential Higher Data Speeds Faster Data Processing Better Data Security Adapters Switches Cables & Transceivers SmartNIC System on a Chip HPC and AI Needs the Most Intelligent Interconnect
- 3. 3© 2018 Mellanox Technologies | Confidential Mellanox Accelerates Leading HPC and AI Systems ‘Summit’ CORAL System World’s Fastest HPC / AI System 9.2K InfiniBand Nodes ‘Sierra’ CORAL System #2 USA Supercomputer 8.6K InfiniBand Nodes Wuxi Supercomputing Center Fastest Supercomputer in China 41K InfiniBand Nodes 1 2 3
- 4. 4© 2018 Mellanox Technologies | Confidential Mellanox Accelerates Leading HPC and AI Systems Fastest Supercomputer in Canada Dragonfly+ Topology 1.5K InfiniBand Nodes ‘Astra’ Arm-Based Supercomputer NNSA Vanguard Program 2.6K InfiniBand Nodes Fastest HPC / AI System in Japan 1.1K InfiniBand Nodes
- 5. 5© 2018 Mellanox Technologies | Confidential Data Centric Architecture to Overcome Latency Bottlenecks CPU-Centric (Onload) Data-Centric (Offload) Communications Latencies of 30-40us Intelligent Interconnect Paves the Road to Exascale Performance GPU CPU GPU CPU GPU CPU CPU GPU GPU CPU GPU CPU GPU CPU CPU GPU Communications Latencies of 3-4us
- 6. 6© 2018 Mellanox Technologies | Confidential In-Network Computing to Enable Data-Centric Data Centers GPU CPU GPU CPU GPU CPU CPU GPU GPUDirect RDMA Scalable Hierarchical Aggregation and Reduction Protocol NVMeOver Fabrics
- 7. 7© 2018 Mellanox Technologies | Confidential SHARP AllReduce Performance Advantages (128 Nodes) SHARP enables 75% Reduction in Latency Providing Scalable Flat Latency
- 8. 8© 2018 Mellanox Technologies | Confidential SHARP AllReduce Performance Advantages 1500 Nodes, 60K MPI Ranks, Dragonfly+ Topology SHARP Enables Highest Performance
- 9. © 2018 UCF Consortium 9 UCXUnified Communication - X Framework WEB: www.openucx.org https://github.com/openucx/ucx Mailing List: https://elist.ornl.gov/mailman/listinfo/ucx-group [email protected]
- 10. 10© 2018 Mellanox Technologies | Confidential Highest-Performance 200Gb/s Interconnect Solutions Transceivers Active Optical and Copper Cables (10 / 25 / 40 / 50 / 56 / 100 / 200Gb/s) 40 HDR (200Gb/s) InfiniBand Ports 80 HDR100 InfiniBand Ports Throughput of 16Tb/s, <90ns Latency 200Gb/s Adapter, 0.6us latency 200 million messages per second (10 / 25 / 40 / 50 / 56 / 100 / 200Gb/s) 16 400GbE, 32 200GbE, 128 25/50GbE Ports (10 / 25 / 40 / 50 / 100 / 200 GbE) Throughput of 6.4Tb/s MPI, SHMEM/PGAS, UPC For Commercial and Open Source Applications Leverages Hardware Accelerations System on Chip and SmartNIC Programmable adapter Smart Offloads
- 11. 11© 2018 Mellanox Technologies | Confidential BlueField System on a Chip ▪ Tile Architecture - 16 ARM® A72 CPUs ▪ SkyMesh™ fully coherent low-latency interconnect ▪ 8MB L2 Cache, 8 Tiles ▪ 12MB L3 Last Level Cache ▪ Integrated ConnectX-5 subsystem ▪ Dual 100Gb/s Ethernet/InfiniBand, compatible with ConnectX-5 ▪ High-end Networking Offloads ▪ RDMA, NVMe-oF, Erasure Coding, T10-DIF ▪ Fully Integrated PCIe switch ▪ 32 Bifurcated PCI Gen4 lanes (up to 200Gb/s) ▪ Root Complex or Endpoint modes ▪ 2x16, 4x8, 8x4 or 16x2 configurations ▪ Crypto Engines ▪ AES, SHA-1/2 ▪ Public Key acceleration, True RNG ▪ Memory Controllers ▪ 2x Channels DDR4 Memory Controllers w/ ECC Dual VPI Ports Ethernet/InfiniBand: 1, 10, 25,40,50,100G 32-lanes PCIe Gen3/4
- 12. 12© 2018 Mellanox Technologies | Confidential BlueField Product Line ▪ Different SKUs ▪ # Cores ▪ Speeds ▪ Perf points SmartNIC Ethernet Dual Port 100Gb/s Controller BF1600 Dual Port 100Gb/s Controller BF1700 BlueField™ Platform BF1100 & BF1200System on Chip ▪ Different SKUs ▪ GPU and SSD ▪ 1U and 2U ▪ Up to 16 SSDs ▪ Different SKUs ▪ 16/8/4 cores ▪ Dual 25Gb/s ▪ Storage card▪ Different SKUs ▪ PCIe x16 ▪ PCIe x32 ▪ SDK and Development Tools ▪ Networking , Security Features ▪ Full NVMe storage capability Full Software Enablement SmartNIC VPI BF1660 ▪ VPI 2-ports 100G ▪ 100Gb/s / EDR ▪ 16 Cores
- 13. 13© 2018 Mellanox Technologies | Confidential Thank You