By Garima Kochhar. HPC Innovation Lab. January 2016.

The Intel Xeon Phi bootable processor (architecture codenamed “Knights Landing” – KNL) is ready for prime time. The HPC Innovation Lab has had access to a few engineering test units, and this blog presents the results of our initial benchmarking study. [We also published our results with Cryo-EM workloads on these systems, and that study is available here.]

The KNL processor is from the Intel Xeon Phi product line but is a bootable processor, i.e., the system does not need another processor in it to power on, just the KNL. Unlike the Xeon Phi coprocessors or the NVIDIA K80 and P100 GPU cards that are housed in a system that has a Xeon processor as well, the KNL is the only processor in the server. This necessitates a new server board design and the PowerEdge C6320p is the Dell EMC platform that supports the KNL line of processors. A C6320p server includes support for one KNL processor and six DDR4 memory DIMMs. The network choices include Mellanox InfiniBand EDR, Intel Omni-Path, or choices of add-in 10GbE Ethernet adapters. The platform has the other standard components you’d expect from the PowerEdge line including a 1GbE LOM, iDRAC and systems management capabilities. Further information on C6320p is available here.

The KNL processor models include 16GB of on-package memory called MCDRAM. The MCDRAM can be used in three modes – memory mode, cache mode or hybrid mode. The 16GB of MCDRAM is visible to the OS as addressable memory and must be addressed explicitly by the application when used in memory mode. In cache mode, the MCDRAM is used as the last level cache of the processor. And in hybrid mode, a portion of the MCDRAM is available as memory and the other portion is used as cache. The default setting is cache mode as this is expected to benefit most applications. This setting is configurable in the server BIOS.

The architecture of the KNL processor allows the processor cores + cache and home agent directory + memory to be organized into different clustering modes. These modes are called all2all, quadrant and hemisphere, Sub-NUMA Clustering-2 and Sub-NUMA Clustering 4. They are described in this Intel article. The default setting in the Dell EMC BIOS is quadrant mode and can be changed in the Dell EMC BIOS. All tests below are with the quadrant mode.

The configuration of the systems used in this study is described in Table 1.

Table 1 - Test configuration


12 * Dell EMC PowerEdge C6320p


Intel Xeon Phi 7230.  64 cores @ 1.3 GHz, AVX base 1.1 GHz.


96 GB at 2400 MT/s [16 GB * 6 DIMMS]


Intel Omni-Path and Mellanox EDR


Operating System

Red Hat Enterprise Linux 7.2


Intel 2017, Build 20160721


Intel MPI 5.1.3


Intel XPPSL 1.4.1


The first check was to measure the memory bandwidth on the KNL system. To measure memory bandwidth to the MCDRAM, the system must be in “memory” mode. A snippet of the OS view when the system is in quadrant + memory mode is in Figure 1.

Note that the system presents two NUMA nodes. One NUMA node contains all the cores (64 cores * 4 logical siblings per physical core) and the 96 GB of DDR4 memory. The second NUMA node, node1, contains the 16GB of MCDRAM.


Figure 1 – NUMA layout in quadrant+memory mode

On this system, the dmidecode command shows six DDR4 memory DIMMs, and eight 2GB MCDRAM memory chips that make up the 16GB MCDRAM.

STREAM Triad results to the MCDRAM on the Intel Xeon Phi 7230 measured between 474-487 GB/s across 16 servers.  The memory bandwidth to the DDR4 memory is between 83-85 GB/s. This is expected performance for this processor model. This link includes information on running stream on KNL.

When the system has MCDRAM in cache mode, the STREAM binary used for DDR4 performance above reports memory bandwidth of 330-345 GB/s.

XPPSL includes a micprun utility that makes it easy to run this micro-benchmark on the MCDRAM. “micprun –k stream –p <num cores>” is the command to run a quick stream test and this will pick the MCDRAM (NUMA node1) automatically if available.


The KNL processor architecture supports AVX512 instructions. With two vector units per core, this allows the processor to execute 32 DP floating point operations per cycle. For the same core count and processor speed, this doubles the floating point capabilities of KNL when compared to Xeon v4 or v3 processors (Broadwell or Haswell) that can do only 16 FLOPS/cycle.

HPL performance on KNL is slightly better (up to 5%) with the MCDRAM in memory mode when compared to cache mode and when using the HPL binary packaged with Intel MKL. Therefore the tests below are with the system in quadrant+memory mode.

On our test systems, we measured between 1.7 – 1.9 TFLOP/s HPL performance per server across 16 test servers. The micprun utility mentioned above is an easy way to run single server HPL tests. “micprun –k hplinpack –p <problem size>” is the command to run a quick HPL test. However for cluster-level tests, the Intel MKL HPL binary is best.

HPL cluster level performance over the Intel Omni-Path interconnect is plotted in Figure 2. These tests were run using the HPL binary that is packaged with Intel MKL. The results with InfiniBand EDR are similar.


Figure 2 - HPL performance over Intel Omni-Path

Conclusion and Next Steps

The KNL-based system is a good platform for highly parallel vector applications. The on-package MCDRAM helps balance the enhanced processing capability with additional memory bandwidth to the applications. KNL introduces the AVX512 instruction set which further improves the performance of vector operations. The PowerEdge C6320p provides a complete HPC server with multiple network choices, disk configurations and systems management.

This blog presents initial system benchmark results. Look for upcoming studies with HPCG and applications like NAMD and Quantum Espresso. We have already published our results with Cryo-EM workloads on KNL and that study is available here.