ThinkSystem Mellanox ConnectX-6 HDR/200GbE VPI Adapters

The ThinkSystem Mellanox ConnectX-6 HDR/200GbE VPI Adapters offer 200 Gb/s Ethernet and InfiniBand connectivity for high-performance connectivity when running HPC, cloud, storage and machine learning applications.

The following figure shows the ThinkSystem Mellanox ConnectX-6 HDR/200GbE QSFP56 1-port PCIe 4 VPI Adapter connected to the ThinkSystem Mellanox HDR/200GbE Aux Adapter (the standard heat sink has been removed in this photo).

Figure 1. ThinkSystem Mellanox ConnectX-6 HDR/200GbE QSFP56 1-port PCIe 4 VPI Adapter (right) and ThinkSystem Mellanox HDR/200GbE Aux Adapter (left)

Mellanox ConnectX-6 brings new acceleration engines for maximizing High Performance, Machine Learning, Storage, Web 2.0, Cloud, Data Analytics and Telecommunications platforms. ConnectX-6 HDR adapters support up to 200 Gb/s total bandwidth at sub-600 nanosecond latency, and NVMe over Fabric offloads, providing the highest performance and most flexible solution for the most demanding applications and markets. ThinkSystem servers with Mellanox adapters and switches deliver the most intelligent fabrics for High Performance Computing clusters.

For servers with support for PCIe 4.0 host interfaces, the ConnectX-6 HDR adapter can be used by itself in a single PCIe 4.0 x16 slot to provide 200 Gb/s connectivity. For servers with PCIe 3.0 interfaces, the ConnectX-6 HDR adapter is used in conjunction with the Aux adapter. The HDR adapter and the Aux adapter are connected together via a cable (included with the Aux adapter) and their combined host interfaces of PCIe 3.0 x32 provides enough bandwidth for 200 Gb/s connectivity.

The following table shows the part numbers for the adapters.

Part number 4C57A15326 includes the following:

• One Mellanox adapter with full-height (3U) adapter bracket attached
• Low-profile (2U) adapter bracket
• Documentation

Part number 4C57A14179 includes the following:

• One Mellanox adapter with full-height (3U) adapter bracket attached
• Low-profile (2U) adapter bracket
• 350mm cable
• Documentation

Note: 4C57A15326 was previously named ThinkSystem Mellanox ConnectX-6 HDR QSFP56 1-port PCIe 4 InfiniBand Adapter

The adapter has an empty QSFP56 cage for connectivity.

The following table lists the supported transceivers.

Configuration notes:

• Transceiver AV1D also supports 40Gb when installed in a Mellanox adapter.
• For the transceiver and cable support for the Mellanox QSA 100G to 25G Cable Adapter (4G17A10853), see the 25G Cable Adapter transceiver and cable support section.

The following table lists the supported fiber optic cables and Active Optical Cables.

25G Cable Adapter transceiver and cable support

The Mellanox QSA 100G to 25G Cable Adapter (4G17A10853) supports the transceivers listed in the following table.

The Mellanox QSA 100G to 25G Cable Adapter (4G17A10853) supports the fiber optic cables and Active Optical Cables listed in the following table.

Machine learning and big data environments

Data analytics has become an essential function within many enterprise data centers, clouds and hyperscale platforms. Machine learning relies on especially high throughput and low latency to train deep neural networks and to improve recognition and classification accuracy. ConnectX-6 offers an excellent solution to provide machine learning applications with the levels of performance and scalability that they require.

ConnectX-6 utilizes the RDMA technology to deliver low-latency and high performance. ConnectX-6 enhances RDMA network capabilities even further by delivering end-to-end packet level flow control.

By performing block-storage encryption in the adapter, ConnectX-6 excludes the need for self-encrypted disks. This allows customers the freedom to choose their preferred storage device, including those that traditionally do not provide encryption. ConnectX-6 can support Federal Information Processing Standards (FIPS) compliance.

ConnectX-6 also includes a hardware Root-of-Trust (RoT), which uses HMAC relying on a device-unique key. This provides both a secure boot as well as cloning-protection. Delivering best-in-class device and firmware protection, ConnectX-6 also provides secured debugging capabilities, without the need for physical access.

Storage environments

NVMe storage devices offer very fast access to storage media. The evolving NVMe over Fabric (NVMe-oF) protocol leverages RDMA connectivity to remotely access NVMe storage devices efficiently, while keeping the end-to-end NVMe model at lowest latency. With its NVMe-oF target and initiator offloads, ConnectX-6 brings further optimization to NVMe-oF, enhancing CPU utilization and scalability.

Cloud and Web 2.0 environments

Telco, Cloud and Web 2.0 customers developing their platforms on software-defined network (SDN) environments are leveraging the Virtual Switching capabilities of server operating systems to enable maximum flexibility in the management and routing protocols of their networks.

Open V-Switch (OVS) is an example of a virtual switch that allows virtual machines to communicate among themselves and with the outside world. Software-based virtual switches, traditionally residing in the hypervisor, are CPU intensive, affecting system performance and preventing full utilization of available CPU for compute functions.

To address such performance issues, ConnectX-6 offers Mellanox Accelerated Switching and Packet Processing (ASAP²) Direct technology. ASAP² offloads the vSwitch/vRouter by handling the data plane in the NIC hardware while maintaining the control plane unmodified. As a result, significantly higher vSwitch/vRouter performance is achieved minus the associated CPU load.

The vSwitch/vRouter offload functions supported by ConnectX-5 and ConnectX-6 include encapsulation and de-capsulation of overlay network headers, as well as stateless offloads of inner packets, packet headers re-write (enabling NAT functionality), hairpin, and more.

In addition, ConnectX-6 offers intelligent flexible pipeline capabilities, including programmable flexible parser and flexible match-action tables, which enable hardware offloads for future protocols.

Socket Direct

Mellanox’s Socket Direct technology improves the performance of dual-socket servers in numerous ways, such as by enabling each of their CPUs to access the network through a dedicated PCIe interface. As the connection from each CPU to the network bypasses the QPI (UPI) and the second CPU, Socket Direct reduces latency and CPU utilization. Moreover, each CPU handles only its own traffic (and not that of the second CPU), thus optimizing CPU utilization even further.

Socket Direct also enables GPUDirect RDMA for all CPU/GPU pairs by ensuring that GPUs are linked to the CPUs closest to the adapter card. Socket Direct enables Intel DDIO optimization on both sockets by creating a direct connection between the sockets and the adapter card.

Socket Direct technology is enabled by a main card housing the ConnectX-6 and an auxiliary PCIe card bringing in the remaining PCIe lanes. The ConnectX-6 Socket Direct card is installed into two PCIe x16 slots and connected using the supplied 350mm cable.

The two PCIe x16 slots may also be connected to the same CPU. In this case, the main advantage of the technology lies in delivering 200Gb/s to servers with PCIe Gen3-only support.

The following figure shows the Mellanox SharedIO Adapter and Auxiliary Card installed in two ThinkSystem SD650 V2 servers in the same tray.

Figure 2. SharedIO adapters installed in the two SD650 V2 servers on a tray

The adapters have the following technical specifications.

Form factor

• Single-slot low-profile main adapter (6.6-inch x 2.71 in.)
• Single-slot low-profile auxilary adapter

PCI Express Interface

• Supports PCIe 4.0 or PCIe 3.0
  • In PCIe 4.0 servers, the ConnectX-6 adapter is used by itself to connect 16 PCIe lanes. For two-socket servers, the adapter can be used with the Aux adapter to enable the Socket Direct feature.
  • In PCIe 3.0 servers, the ConnectX-6 adapter is used with the Aux adapter to connect 32 PCIe lanes. For two-socket servers, the adapter can be used with the Aux adapter to enable the Socket Direct feature.
• PCIe Atomic
• TLP (Transaction Layer Packet) Processing Hints (TPH)
• PCIe switch Downstream Port Containment (DPC) enablement for PCIe hot-plug
• Advanced Error Reporting (AER)
• Access Control Service (ACS) for peer-to-peer secure communication
• Process Address Space ID (PASID) Address Translation Services (ATS)
• IBM CAPIv2 (Coherent Accelerator Processor Interface)
• Support for MSI/MSI-X mechanisms

Connectivity

• One QSFP56 port
• Supports passive copper cables with ESD protection
• Powered connectors for optical and active cable support

InfiniBand

• Supports interoperability with InfiniBand switches (up to HDR, as 4 lanes of 50Gb/s data rate)
• Total connectivity is 200 Gb/s:
  • One port adapter supports a single 200 Gb/s link
• HDR / HDR100 / EDR / FDR / QDR / DDR / SDR
• IBTA Specification 1.3 compliant
• RDMA, Send/Receive semantics
• Hardware-based congestion control
• Atomic operations
• 16 million I/O channels
• 256 to 4Kbyte MTU, 2Gbyte messages
• 8 virtual lanes + VL15

Enhanced Features

• Hardware-based reliable transport
• Collective operations offloads
• Vector collective operations offloads
• PeerDirect RDMA (GPUDirect) communication acceleration
• 64/66 encoding
• Enhanced Atomic operations
• Advanced memory mapping support, allowing user mode registration and remapping of memory (UMR)
• Extended Reliable Connected transport (XRC)
• Dynamically Connected transport (DCT)
• On demand paging (ODP)
• MPI Tag Matching
• Rendezvous protocol offload
• Out-of-order RDMA supporting Adaptive Routing
• Burst buffer offload
• In-Network Memory registration-free RDMA memory access

CPU Offloads

• RDMA over Converged Ethernet (RoCE)
• TCP/UDP/IP stateless offload
• LSO, LRO, checksum offload
• RSS (also on encapsulated packet), TSS, HDS, VLAN and MPLS tag insertion/stripping, Receive flow steering
• Data Plane Development Kit (DPDK) for kernel bypass applications
• Open VSwitch (OVS) offload using ASAP²
  • Flexible match-action flow tables
  • Tunneling encapsulation / de-capsulation
• Intelligent interrupt coalescence
• Header rewrite supporting hardware offload of NAT router

Storage Offloads

• Block-level encryption: XTS-AES 256/512 bit key
• NVMe over Fabric offloads for target machine
• Erasure Coding offload - offloading Reed-Solomon calculations
• T10 DIF - signature handover operation at wire speed, for ingress and egress traffic
• Storage Protocols: SRP, iSER, NFS RDMA, SMB Direct, NVMe-oF

Overlay Networks

• RoCE over overlay networks
• Stateless offloads for overlay network tunneling protocols
• Hardware offload of encapsulation and decapsulation of VXLAN, NVGRE, and GENEVE overlay networks

Hardware-Based I/O Virtualization

• Single Root IOV
• Address translation and protection
• VMware NetQueue support
• SR-IOV: Up to 512 Virtual Functions
• SR-IOV: Up to 16 Physical Functions per host
• Virtualization hierarchies (network partitioning, NPAR)
  • Virtualizing Physical Functions on a physical port
  • SR-IOV on every Physical Function
• Configurable and user-programmable QoS
• Guaranteed QoS for VMs

Management and Control

• NC-SI, MCTP over SMBus and MCTP over PCIe - BMC interface
• PLDM for Monitor and Control DSP0248
• PLDM for Firmware Update DSP0267
• SDN management interface for managing the eSwitch
• I2C interface for device control and configuration
• General Purpose I/O pins
• SPI interface to Flash
• JTAG IEEE 1149.1 and IEEE 1149.6

NVIDIA Unified Fabric Manager (UFM) is InfiniBand networking management software that combines enhanced, real-time network telemetry with fabric visibility and control to support scale-out InfiniBand data centers.

The two offerings available from Lenovo are as follows:

• UFM Telemetry for Real-Time Monitoring

  The UFM Telemetry platform provides network validation tools to monitor network performance and conditions, capturing and streaming rich real-time network telemetry information, application workload usage, and system configuration to an on-premises or cloud-based database for further analysis.

• UFM Enterprise for Fabric Visibility and Control

  The UFM Enterprise platform combines the benefits of UFM Telemetry with enhanced network monitoring and management. It performs automated network discovery and provisioning, traffic monitoring, and congestion discovery. It also enables job schedule provisioning and integrates with industry-leading job schedulers and cloud and cluster managers, including Slurm and Platform Load Sharing Facility (LSF).

The following table lists the subscription licenses available from Lenovo.

For more information, see the following web page:
https://www.nvidia.com/en-us/networking/infiniband/ufm/

The following servers offer a PCIe 4.0 host interface. All other supported servers have a PCIe 3.0 host interface.

• ThinkSystem SR635
• ThinkSystem SR655
• ThinkSystem SR645
• ThinkSystem SR665

The following tables list the ThinkSystem servers that are compatible.

The adapters support the operating systems listed in the following tables.

• ThinkSystem Mellanox ConnectX-6 HDR/200GbE QSFP56 1-port PCIe 4 VPI Adapter, 4C57A15326
• ThinkSystem Mellanox HDR/200GbE 2x PCIe Aux Kit, 4C57A14179

The adapters have the following regulatory approvals:

• Safety: CB / cTUVus / CE
• EMC: CE / FCC / VCCI / ICES / RCM / KC
• RoHS: RoHS Compliant

The adapters have the following operating characteristics:

• Typical power consumption (passive cables): 19.3W
• Maximum power available through QSFP56 port: 5W
• Temperature
  • Operational: 0°C to 55°C
  • Non-operational: -40°C to 70°C
• Humidity: 90% relative humidity

One year limited warranty. When installed in a Lenovo server, the adapter assumes the server’s base warranty and any warranty upgrades.

For more information, refer to these documents:

• Networking Options for ThinkSystem Servers:
  https://lenovopress.com/lp0765-networking-options-for-thinksystem-servers
• ServerProven compatibility:
  http://www.lenovo.com/us/en/serverproven
• Mellanox InfiniBand product page:
  https://www.nvidia.com/en-us/networking/infiniband-adapters/
• ConnectX-6 VPI user manual:
  https://docs.nvidia.com/networking/display/ConnectX6VPI