Intel Omni-Path
Intel Omni-Path is a computer networking technology that competes with InfiniBand . The abbreviation for the product family is OPA (Intel Omni-Path Architecture). This is a manufacturer-specific technology and not an open standard like Infiniband. This technology is used in the area of "high performance computing". Intel thus supplies all the individual components for the technology: host bus adapter, CPU adapter, cables and switches; however, the components are mainly sold through partners (supercomputer manufacturers).
Standardized optical cables with QSFP 28 connectors are used. When it was released in November 2015, Omni-Path supports a maximum bandwidth of 100 GBit / s and is therefore on par with Infiniband EDR.
The interface is supported with the processors of the following series:
The communication protocol is based closely on InfiniBand, but has three main changes:
- Link Transfer Packet : Easier error correction in the link transfer layer, reduces latency with low error rates
- Traffic Flow Optimization : packing several messages into one package
- Quality of Service : messages with a higher priority can interrupt ongoing transmissions
The following Intel CPUs are available with a connector for Omni-Path integrated into the processor:
PCIe 4/8 lanes (Generation 3.0) host bus adapters for Intel Omni-Path are also available for other processors.
With the Xeon SP, Intel apparently uses the connections of the processors for a QPI / UPI link for the Omni-Path interface, Xeon SP processors with an Omni-Path interface only have 2 UPI links. The OPA interface integrated in the processor is also connected via PCIe.
In August 2019, Intel announced that it would be discontinuing the further development OPA-200. Older OPA-100 with 100 Gbps and multi-railing are still supported.
Omni-Path in the TOP500 supercomputers
| date | Number of systems |
|---|---|
| June 2016 | 8th |
| November 2016 | 20th |
| June 2017 | 38 |
| November 2017 | 51 |
Since the release date (11/2015), Intel has had a number of installed systems with omni-path technology in the supercomputer lists, according to its own information about 50% of the newly installed clusters.
The largest installation of an Omni-Path cluster on the June 2017 list is the Oakforest PACS , in 7th place in the TOP500 and the fastest system in Japan, based on Intel Xeon Phi processors with a total of 8,208 compute nodes.
Web links
- Crosslinking products
- Omni-Path Switches
- Omni-Path PCI-E adapter
- Omni-Path Cable
- A New High Performance Computing Fabric White Paper
Individual evidence
- ↑ Omni-Path page of the Niflheim cluster. Retrieved September 12, 2017 .
- ↑ Intel Omni-Path Details. Retrieved September 6, 2017 .
- ↑ exploring Intel's Omni-Path. Retrieved September 6, 2017 .
- ↑ Intel Skylake Omni-Path Fabric Does Not Work on Every Server and Motherboard. Retrieved September 29, 2017 .
- ↑ Intel stops second generation of the fast InterConnect Omni-Path. Retrieved January 24, 2020 .
- ↑ Intel Omni-Path Updates at SC16 (English). Retrieved September 12, 2017 .
- ↑ Oakforest PACS in the TOP500 (English). Retrieved September 29, 2017 .
- ↑ How Oakforest-PACS Outpaced The K Supercomputer (English). Retrieved September 29, 2017 .
- ^ Supercomputer lists