White Papers

ManualsBrandsDell ManualsConverged InfrastructureHigh Performance Computing Solution Resources

Ready Solutions Engineering Test Results

14G with Skylake – how much better for HPC?

Dell EMC HPC Innovation Lab, September 2017

This document describes the features of the latest Dell EMC servers and compares performance to previous generation systems for a

variety of HPC applications.

New servers and Skylake

Dell EMC recently announced the 14

generation PowerEdge server portfolio (14G) which supports the latest generation Intel Scalable

Processor Family (the micro-architecture that is code named “Skylake”). In addition to the latest processor and accelerator/GPU

support, 14G includes several other technology additions. These servers have enhanced systems management and security features

via iDRAC9, can support up to 24 NVMe drives per server, include support for NVDIMM and future 3D XPoint memory technologies,

allow options for direct contact liquid cooling within the server, etc.

This document is focused on the performance gains available with the latest generation Intel Skylake CPUs in Dell EMC 14G platforms.

The Skylake CPU (SKL) supports six DDR4 memory channels per socket with memory modules that can run at up to 2667 MT/s. SKL

provides up to 28 cores per socket with TDP up to 205 W per socket. Intel has introduced AVX512 instructions on SKL, and this

doubles the floating point capabilities of SKL over previous generation Xeons. The processor now supports 512 bit registers, and the

Platinum 8100 and Gold 6100 SKL CPU models have two fuse multiply add (FMA) units, each of which can execute 8 double precision

calculations per cycle. With two floating point operations per FMA instruction, Skylake can execute 32 FLOP/cycle, double the previous

generation Xeon which was 16 FLOP/cycle. Note that some models of the SKL CPU like the Gold 5100, Silver 4100 and Bronze 3100

CPUs have one FMA unit, giving 16 FLOP/cycle. As before, the CPU frequency will depend on whether the code has a high density of

AVX2 or AVX512 instructions. An application will run at faster CPU clock speeds when running non-AVX codes, and will run slowest

with a high density of AVX512 instructions.

Other changes in SKL include 48 PCIe lanes per socket vs. 40 lanes previously, and a new interconnect called UPI (Ultra Path

Interconnect) between the sockets, replacing the previous QPI interconnect. UPI can operate at up to 10.4 GT/s, faster than the 9.6

GT/s with QPI. There are other architectural changes like a larger L2 cache for the cores, a non-inclusive L3 cache, a new uncore

interconnect, distributed home agent, optimized turbo bins, per core P-states, etc. Architectural changes in the silicon lead to new

tuning options in the BIOS and one of these, Sub NUMA Clustering, is discussed in this blog.

We focus on measuring full system performance and compare 14G compute centric performance to the previous generation Dell EMC

platforms. Some of the performance improvements are due to faster memory, some due to AVX512, some due to additional cores and

some due to the combination of all the Intel micro-architecture enhancements. We show results for up to six generations of Intel

processors and four generations of Dell EMC servers. The storage system and I/O is not a significant portion in these tests.

The shorthand used in the graphs below is explained here.

o 11G WSM – Dell EMC 11

generation servers with support for Intel Xeon 5600 series processors, micro-architecture

code named Westmere (WSM).

o 12G SB – Dell EMC 12

generation servers with support for Intel Xeon 2600 series processors, micro-architecture code

named Sandy Bridge (SB).

o 12G IVB – Dell EMC 12

generation servers with support for Intel Xeon 2600 v2 series processors, micro-architecture

code named Ivy Bridge (IVB).

o 13G HSW – Dell EMC 13

generation servers with support for Intel Xeon 2600 v3 series processors, micro-architecture

code named Haswell (HSW).

o 13G BDW – Dell EMC 13

generation servers with support for Intel Xeon 2600 v4 series processors, code named

Broadwell (BDW).

o 14G SKL – Dell EMC 14

generation servers with support for Intel Xeon Scalable Processor Family, micro-architecture

code named Skylake (SKL).

Summary of content (8 pages)

PAGE 1
Ready Solutions Engineering Test Results 14G with Skylake – how much better for HPC? Dell EMC HPC Innovation Lab, September 2017 This document describes the features of the latest Dell EMC servers and compares performance to previous generation systems for a variety of HPC applications.
PAGE 2
Ready Solutions Engineering Test Results 2 System benchmark performance – STREAM and HPL We start our performance study with two standard system benchmarks, STREAM to measure the system memory bandwidth and HPL to measure compute performance. Figure 1 shows the memory bandwidth capability of a two socket (2S) Skylake based system for two CPU models, the mid-tier Gold 6150 and the highest end Platinum 8180.
PAGE 3
Ready Solutions Engineering Test Results 3 Figure 2 - HPL performance over generations Moving on to HPL, Figure 2 shows the compute performance of a server over the last six generations. Similar to the Stream results, data is plotted from Intel Xeon X5600 series (Westmere) to Intel Scalable Processor Family (Skylake) and shows the performance of the system when all cores are in use. The bars show HPL performance for a 2S system and the squares show performance relative to BDW.
PAGE 4
Ready Solutions Engineering Test Results Figure 3 - WRF performance over generations Life Sciences – BWA Pipeline Alignment of raw reads to a reference genome is one of the key steps in next generation sequencing data analysis, and BWA is a software package for mapping low-divergent sequences against a large reference genome, such as the human genome. BWA performance for 416 million reads sequencing data is shown in Figure 4.
PAGE 5
Ready Solutions Engineering Test Results Figure 4 - BWA performance over generations. The height of the bar represents relative performances against a single core of E5 2697 v4. Figure 5 - BWA throughput over generations Copyright © 2017 Dell Inc. or its subsidiaries. All Rights Reserved. Dell, EMC, and other trademarks are trademarks of Dell Inc. or its subsidiaries. Other trademarks may be the property of their respective owners. Published in the USA.
PAGE 6
Ready Solutions Engineering Test Results 6 Manufacturing applications – ANSYS Fluent, STAR-CCM+ and LS-DYNA ANSYS® Fluent® and STAR-CCM+® software are computational fluid dynamics applications commonly used for research and product development. LS-DYNA® is a general-purpose finite element software in use by multiple industries. The system performance for these three applications, across four processor generations, is presented here. Standard benchmark datasets are used for the performance comparisons.
PAGE 7
Ready Solutions Engineering Test Results Figure 7 - STAR-CCM+ performance over generations Figure 8 - LS-DYNA performance over generations Copyright © 2017 Dell Inc. or its subsidiaries. All Rights Reserved. Dell, EMC, and other trademarks are trademarks of Dell Inc. or its subsidiaries. Other trademarks may be the property of their respective owners. Published in the USA. Dell EMC believes the information in this document is accurate as of its publication date.
PAGE 8
Ready Solutions Engineering Test Results 8 The performance for LS-DYNA® for the Car2Car dataset across four processor generations is plotted in Figure 8. In this figure, the height of the bars and the label represents the 2S system performance relative to the 13G Intel Xeon E5-2697A v4 result. The LSDYNA AVX2 binary was used for Haswell through Skylake results, while the SSE2 binary was used for Ivy Bridge, since Ivy Bridge does not support the AVX2 instruction set.