Intel rolled out the Knights Landing, its next-generation Xeon Phi multi-processor chip with three times the performance of its predecessor at 3TFLOPS. The Xeon Phi Knights Landing uses a 2D on-chip mesh interconnect and touts Intel's new Omni-Path off-chip communications.

Until now the Xeon Phi has been mostly used in supercomputers running naturally parallel applications, but many new features in the Knights Landing version may appeal to users running engineering and enterprise workloads. For instance, it is now a stand-alone chip and does not have the need for complex software running on a regular Xeon to supervise it.

The Xeon Phi Knights Landing will house more than 60 cores (rumoured up to 72) intimately married to a special version of Micron's Hybrid Memory Cubes mounted in-package yielding effectively 16GB of on-chip memory. It has enough in-package memory and off-chip DDR (384GB) to run workloads without adding fancy swapping and caching techniques.

In addition, Intel announced that its new Omni-Path alternative to Infiniband (for connecting rack systems) will be an option on its latest Xeon Phi. Omni-Path offers 100GB/s line speed and up to 56 per cent lower switch fabric latency over InfiniBand. The Omni-Path Architecture will also use a 48 port switch chip to deliver greater port density over InfiniBand's 36 ports.

As a result, Omni-Path offers up to 1.3x greater port density than InfiniBand, uses up to 50 per cent fewer switches than a Infiniband and up to 2.3x higher scaling. The Xeon Phi Knights Landing also features 36 PCIe 3.0 lanes for communicating with peripherals.

Intel also announced the successor to the Xeon Phi Knights Landing—called Knight's Hill—which will be cast in Intel's 10nm process and have Omni-Path built-in on every chip.

By: DocMemory
Copyright © 2023 CST, Inc. All Rights Reserved