Intel Offers More Details of Its Omni-Path Interconnect Fabric
Intel also is working on Omni-Path PCIe cards with one or two ports that can be used with chips that are not Xeon Phis, and is developing a 48-port switch that will feature Omni-Path. Overall, the Omni-Path Architecture will be 25 percent to 40 percent less expensive than InfiniBand, and will be able to be used in data centers as well as HPC environments, something that doesn't tend to happen with InfiniBand, according to Murphy. In addition, it will offer 73 percent higher switch MPI message rate and 33 percent lower latency than other interconnect fabrics. The Intel officials noted that one innovation that will help reduce latency is the insertion of a new networking layer between the traditional Layers 1 and 2. Murphy called it Layer 1.5—the link transport layer—which essentially reduces workload packets into smaller 65-bit units called "flits." The link transport layer initially was being worked on at Cray, and came to Intel after the chip maker bought the networking technology. A set of 16 flits plus the CRC (Cyclic Redundancy Check) create what officials call a packet. The technology reduces latency by enabling high-priority packets to move larger and lower-priority packets, even if those larger packets have already been sent down the line, Murphy said. This ensures traffic optimization so that the high-priority packets get to their destinations as fast as possible.So far, the interest from systems makers has been good, according to Saleh. There are more than 100 OEM system designs in the works that will feature Omni-Path Architecture, and more than 100,000 nodes are under bid or contract. The first deployments are expected in the fourth quarter, with momentum continuing into next year, he said.
The Intel officials also stressed the importance of open-source technology to the Omni-Path Architecture. The chip maker is supporting Open Fabric Interfaces from by the open-source group OpenFabrics Alliance, which the officials said will lead to improved performance and the ability to scale to tens of thousands of nodes, Murphy said.