Move ǀ store ǀ process

Transcription

1 Move ǀ store ǀ process

2 EXECUTIVE VICE PRESIDENT & GENERAL MANAGER DATA CENTER GROUP

3 OVER of the World s data WAS CREATED IN THE LAST LESS THAN Has Been Analyzed

4 PROLIFERATION OF GROWTH OF CLOUDIFICATION OF THE

5 AI Analytics HPC Multi-cloud & Orchestration COMPUTE DEMAND Network Database Virtualization Security

6 Move Store Process ETHERNET SILICON PHOTONICS OMNI-PATH FABRIC Software & System-Level

7 L A U N C H I N G T O D A Y MOVE STORE PROCESS

8

9 2 ND GENERATION STANDARD SKUS CUSTOM SKUS CORES PER SOCKET MEMORY PER SOCKET SOCKETS Intel Optane DC Persistent Memory Intel Deep Learning Boost intel Speed Select Technology network-optimized SKUs Cloud-optimized SKUs Security Mitigations BUILDING ON 20 YEARS OF DATA CENTER PROCESSOR INNOVATION

10 All New level of Advanced Performance up to 56 cores up to 12 channels PER SOCKET performance per rack DESIGNED FOR THE MOST DATA-INTENSIVE WORKLOADS NATIVE DDR4 MEMORY CPU CPU PACKAGE

11

12 BUSINESS ANALYTICS BUSINESS ANALYTICS 9242 VS VS 8160 BUSINESS ANALYTICS 8280+OPTANE PM VS DRAM BUSINESS ANALYTICS 8280+OPTANE PM VS DRAM maximizing mainstream SKUs up to average perf gain gen on gen CLOUD MANAGEMENT IN-MEMORY DATABASE MORE VMS 8260+OPTANE PM VS DRAM 8260+OPTANE PM VS DRAM BUSINESS ANALYTICS LOWER LATENCY 8260 DLBOOST VS FP32 BUSINESS ANALYTICS 8260 DLBOOST VS FP32 VNETWORK GATEWAY 5218N+QAT VS 5118

13 AI DATA CENTER LOGIC SILICON TAM Inference INFERENCE of the AI si Opportunity TRAINING

14 INFERENCE THROUGHPUT (IMAGES/SEC) OPTIMIZING AI INFERENCE INTEL OPTIMIZATION FOR CAFFE RESNET-50 intel xeon platinum intel xeon platinum INTEL AVX-512 INTEL DL BOOST JUL'17 DEC'18 APR'19 COLUMN1 BASE VS BASE VS BASE intel xeon platinum 8100 processor

15 I N T E L D L B O O S T E C O S Y S T E M S U P P O R T IMAGE RECOGNITION VIDEO ANALYSIS TEXT DETECTION 8 DIFFERENT WORKLOADS ML INFERENCING OPTIMIZED SW & FRAMEWORKS SOFTWARE VENDORS CLOUD SERVICE PROVIDERS ENTERPRISES

16 VICE PRESIDENT AWS COMPUTE SERVICES

17 Data Center Cloud Core Access Edge Devices Things DEFINED PROOF OF CONCEPTS OF COMMS SPS ADOPT NFV MOVES TO LINUX FOUNDATION NETWORK IS VIRTUALIZED CLOUD-NATIVE NETWORK

18 2 nd Gen Intel Xeon Scalable Processors with Intel speed select Technology up to Network Workload Performance VS 1 ST GENERATION INTEL XEON SCALABLE

19

20

21 @mattbytes SR. STAFF HARDWARE ENGINEER TWITTER

22 L A U N C H I N G T O D A Y MOVE STORE PROCESS

23 FOR MISSION CRITICAL ENTERPRISE STORAGE up to STORAGE RACK CONSOLIDATION VS HARD DRIVES

24 MEMORY INNOVATION 10 YEARS IN THE MAKING ECOSYSTEM SUPPORT up to 8 SOCKET SYSTEM BW on HANA Records NEW WORLD RECORD SOLUTION OPTIMIZATION TECHNOLOGY INNOVATIONS up more VM instances to MEETING SUB-mS SLA

25 SENIOR VICE PRESIDENT HEAD OF DATABASE

26

27 VICE PRESIDENT PLATFORMS

28 L A U N C H I N G T O D A Y MOVE STORE PROCESS

29

30 System Configuration: Leadership performance per rack Intel's compilers may or may not optimize to the same degree for non-intel microprocessors for optimizations that are not unique to Intel microprocessors. These optimizations include SSE2, SSE3, and SSSE3 instruction sets and other optimizations. Intel does not guarantee the availability, functionality, or effectiveness of any optimization on microprocessors not manufactured by Intel. Microprocessor-dependent optimizations in this product are intended for use with Intel microprocessors. Certain optimizations not specific to Intel microarchitecture are reserved for Intel microprocessors. Please refer to the applicable product User and Reference Guides for more information regarding the specific instruction sets covered by this notice. Performance per rack leadership based on 4 benchmarks (Integer Throughput, Floating Point Throughput, Memory Bandwidth and LINPACK). Details below: Integer Throughput: 1-node, 2x Intel Xeon Platinum 9282 processor on Walker Pass with 768 GB (24x 32GB 2933) total memory, ucode 0x on CentOS Linux release , , IC19u1, AVX512, HT on, Turbo on, score: est int throughput=628, test by Intel on 3/14/2019. Rack performance estimate of U rack, 32U dedicated to compute, total of 64 compute nodes. 64 * 628 = node, 2x AMD* EPYC* 7601, score: 301, test by Dell on Feb 2019 Rack performance estimate of U rack, 32U dedicated to compute, total of 64 compute nodes. 64 * 301 = Floating Point Throughput: 1-node, 2x Intel Xeon Platinum 9282 processor on Walker Pass with 768 GB (24x 32GB 2933) total memory, ucode 0x on CentOS Linux release , , IC19u1, AVX512, HT on, Turbo on, score: est fp throughput=522, test by Intel on 3/14/2019. Rack performance estimate of U rack, 32U dedicated to compute, total of 64 compute nodes. 64 * 522 = node, 2x AMD* EPYC* 7601, score: 282, test by Dell on Feb 2019 Rack performance estimate of U rack, 32U dedicated to compute, total of 64 compute nodes. 64 * 282 = Memory Bandwidth: 1-node, 2x Intel Xeon Platinum 9282 processor on Walker Pass with 768 GB (24x 32GB 2933) total memory, ucode 0x , on CentOS Linux release , , IC19u1, AVX512, HT off, Turbo on, score: Stream Triad=407GiB/s, test by Intel on 3/14/2019. Rack performance estimate of U rack, 32U dedicated to compute, total of 64 compute nodes. 64 * 407 = node, 2x AMD* EPYC* 7601, score=290, test by AMD as of June Rack performance estimate of U rack, 32U dedicated to compute, total of 64 compute nodes. 64 * 290 = LINPACK: 1-node, 2x Intel Xeon Platinum 9282 processor on Walker Pass with 768 GB (24x 32GB 2933) total memory, ucode 0x on CentOS Linux release , , IC19u1, N , AVX512 MKL 2019, HT off, Turbo on, score: Intel Distribution of LINPACK=6411, test by Intel on 3/14/2019. Rack performance estimate of TFlops. 42U rack, 32U dedicated to compute, total of 64 compute nodes. 64 * 6411 = TFlops 1-node, 2x AMD EPYC 7601: Supermicro AS-2023US-TR4 with 2 AMD EPYC 7601 (2.2GHz, 32 core) processors, SMT OFF, Turbo ON, BIOS ver 1.1a, 4/26/2018, microcode: 0x , 16x32GB DDR4-2666, 1 SSD, Ubuntu LTS ( generic Retpoline), High Performance Linpack v2.2, compiled with Intel(R) Parallel Studio XE 2018 for Linux, Intel MPI version , AMD BLIS ver 0.4.0, Benchmark Config: Nb=232, N=168960, P=4, Q=4, Score =1095GFs, tested by Intel as of July 31, Rack performance estimate of TFlops. 42U rack, 32U dedicated to compute, total of 64 compute nodes. 64 * 1095 = TFLOPs

31 System Configuration: World Record + Real Workload Performance Leadership Performance results are based on testing as of dates shown in configuration and may not reflect all publicly available security updates. See configuration disclosure for details. No product or component can be absolutely secure. Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors. Performance tests, such as SYSmark and MobileMark, are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products. For more complete information visit x LAMMPS* Water: 1-node, 2x Intel Xeon Platinum 8160L cpu on Wolf Pass with 192 GB (12 slots / 16GB / 2666) total memory, ucode 0x200004d on Oracle Linux Server release 7.6, el7.crt1.x86_64, Intel SSDSC2BA80, LS-Dyna 9.3- Explicit AVX2 binary, 3car, HT on, Turbo on, test by Intel on 2/26/ node, 2x Intel Xeon Platinum 9242 cpu on Intel reference platform with 384 GB (24 slots / 16GB / 2933) total memory, ucode 0x on CentOS 7.6, el7.x86_64, Intel SSDSC2BA80, LS-Dyna 9.3-Explicit AVX2 binary, 3car, HT on, Turbo on, test by Intel on 3/18/ x LS-Dyna* Explicit, 3car: 1-node, 2x Intel Xeon Platinum 8160L cpu on Wolf Pass with 192 GB (12 slots / 16GB / 2666) total memory, ucode 0x200004d on Oracle Linux Server release 7.6, el7.crt1.x86_64, Intel SSDSC2BA80, LAMMPS version 12 Dec 2018, Water, HT on, Turbo on, test by Intel on 2/26/ node, 2x Intel Xeon Platinum 9242 cpu on Intel reference platform with 384 GB (24 slots / 16GB / 2933) total memory, ucode 0x on CentOS 7.6, el7.x86_64, Intel SSDSC2BA80, LAMMPS version 12 Dec 2018, Water, HT on, Turbo on, test by Intel on 3/8/ x BAOSIGHT* xinsight*: 1-node, 2x Intel Xeon Platinum 8260L cpu on S2600WFS with 768 DDR GB (24 slots / 32GB / 2666) total memory, ucode 0x400000A on CentOS 7.5, el7.x86_64, 1x Intel 480GB SSD OS Drive, 1x Intel XC722, xinsight 2.0 internal workload, HT on, Turbo on, test by Intel/Baosight on 1/8/ node, 2x Intel Xeon Platinum 8260L cpu on S2600WFS with 192 DDR Intel DCPMM GB (12 slots / 16 GB / 2666 DDR + 8 slots / 128 GB / 2666 Intel DCPMM) total memory, ucode 0x400000A on CentOS 7.5, el7.x86_64, 1x Intel 480GB SSD OS Drive, 1x Intel XC722, xinsight 2.0 internal workload, HT on, Turbo on, test by Intel/Baosight on 1/9/ x AsiaInfo* BSS*: 1-node, 2x Intel Xeon Platinum 8180 cpu on S2600WFD with 768 GB (24 slots / 32GB / 2666) total memory, ucode 0x on RedHat 7.5, el7.x86_64, 1x Intel 400GB SSD OS Drive, 1 x P4500 1TB Application Data, 1x Intel XC722, BSS self defined workload, HT on, Turbo on, test by Intel/AsiaInfo on 12/27/ node, 2x Intel Xeon Platinum 8280 cpu on S2600WFD with 192 DDR Intel DCPMM GB (12 slots / 16 GB / 2666 DDR + 8 slots / 128 GB / 2666 Intel DCPMM) total memory, ucode 0x400000A on RedHat 7.5, el7.x86_64, 1x Intel 400GB SSD OS Drive,, 1x Intel XC722, BSS self defined workload, HT on, Turbo on, test by Intel/AsiaInfo on 12/26/ x Huawei* FusionSphere*: 1-node, 2x Intel Xeon Platinum 8260L cpu on Wolf Pass with 1024 GB (16 slots / 64GB / 2666) total memory, ucode 0x400000A on FusionSphere HyperV, _96.x86_64, 1x Intel 800GB SSD OS Drive, 1x Intel 800GB SSD OS Drive, 1x Intel XC722, FusionSphere 6.3.1, mysql , sysbench-1.0.6, HT on, Turbo on, test by Huawei/Intel on 1/11/ node, 2x Intel Xeon Platinum 8260L cpu on Wolf Pass with 384 DDR Intel DCPMM GB (12 slots / 32 GB / 2666 DDR + 12 slots / 128 GB / 2666 Intel DCPMM) total memory, ucode 0x400000A on FusionSphere HyperV, _96.x86_64, 3 x P TB Application Data, 3 x P TB Application Data, 1x Intel XC722, FusionSphere 6.3.1, mysql , sysbench-1.0.6, HT on, Turbo on, test by Huawei/Intel on 1/11/ x GBASE: 1-node, 2x Intel Xeon Platinum 8260 cpu on S2600WFT with 768 DDR GB (24 slots / 32GB / 2666) total memory, ucode 0x400000A on CentOS 7.5, el7.x86_64, 1x Intel 400GB SSD OS Drive, 1x Intel XC722, Gbase 8m OCS Benchmark, HT on, Turbo on, test by GBASE/Intel on 2/19/ node, 2x Intel Xeon Platinum 8260 cpu on S2600WFT with 192 DDR Intel DCPMM GB (12 slots / 16 GB / 2666 DDR + 8 slots / 128 GB / 2666 Intel DCPMM) total memory, ucode 0x400000A on CentOS 7.5, el7.x86_64, 1x Intel 400GB SSD OS Drive, 1x Intel XC722, Gbase 8m OCS Benchmark, HT on, Turbo on, test by GBASE/Intel on 2/19/2019. Up to 1.33x average generational gains on mainstream Gold SKU: Geomean of est SPECrate2017_int_base, est SPECrate2017_fp_base, Stream Triad, Intel Distribution of Linpack, server side Java. Gold 5218 vs Gold 5118: 1-node, 2x Intel Xeon Gold 5218 cpu on Wolf Pass with 384 GB (12 X 32GB 2933 (2666)) total memory, ucode 0x on RHEL7.6, el7.x86_64, IC18u2, AVX2, HT on all (off Stream, Linpack), Turbo on, result: est int throughput=162, est fp throughput=172, Stream Triad=185, Linpack=1088, server side java=98333, test by Intel on 12/7/ node, 2x Intel Xeon Gold 5118 cpu on Wolf Pass with 384 GB (12 X 32GB 2666 (2400)) total memory, ucode 0x200004D on RHEL7.6, el7.x86_64, IC18u2, AVX2, HT on all (off Stream, Linpack), Turbo on, result: est int throughput=119, est fp throughput=134, Stream Triad=148.6, Linpack=822, server side java=67434, test by Intel on 11/12/ x Cloudwalk inference latency improvement: 1-node, 2x Intel Xeon Platinum 8260L cpu on S2600WFS with 192 GB (12 slots / 16 GB / 2666 MHz) total memory, ucode 0x400000A on CentOS 7.5, el7.x86_64, 1x Intel 480GB SSD OS Drive, 1 x P4500 1TB Application Data, 1x Intel XC722, Cloudwalk Facial Recognition, GCC 4.8.5, Intel MKL-DNN, Intel Optimization for Caffe 1.1.2, Custom ResNet50, HT on, Turbo on, Comparing inference latency performance on same system with FP32 vs INT8 w/ Intel DL Boost, test by Cloudwalk/Intel on 2/15/ x face recognition performance improvement for HiSign: Tested by Intel and HiSign as of 02/01/ socket Intel Xeon Platinum 8260 Processor, 24 cores HT On Turbo ON Total Memory 768 GB (12 slots/ 64GB/ 2666 MHz), BIOS version (ucode 0x400000A), RedHat 7.5 kernel el7.elrepo.x86_64, Compiler: gcc 4.8.5, Deep Learning Framework: Intel Optimizations for Caffe v1.1.2, Topology: modified Resnet32,custom dataset, BS=1. Comparing performance on same system with FP32 vs INT8 w/ Intel DL Boost 2x Nokia* SDWAN: Configuration #1 (With Intel QuickAssist Technology): 2x Intel Xeon Gold 5218N Processor on Neon City Platform with 192 GB total memory (12 slots / 16GB / DDR4 2667MHz), ucode 0x , Bios: PLYXCRB 1.86B.0568.D , ucode: 0x on CentOS 7.5 with Kernel , KVM Hypervisor; 1x Intel QuickAssist Adapter 8970, Cipher: AES-128 SHA-256; Intel Ethernet Converged Network Adapter X520-SR2; Application: Nokia Nuage SDWAN NSGv 5.3.3U3. Configuration # 2: 2x Intel Xeon Gold 5118 Processor on Neon City Platform with 192 GB total memory (12 slots / 16GB / DDR4 2667MHz), ucode 0x , Bios: PLYXCRB 1.86B.0568.D , ucode: 0x on CentOS 7.5 with Kernel , KVM Hypervisor; Intel Ethernet Converged Network Adapter X520-SR2; Application: Nokia Nuage SDWAN NSGv 5.3.3U3. Results recorded by Intel on 2/14/2018 in collaborate with Nokia.

32 System Configuration: Intel Deep Learning Boost 1x inference throughput baseline on Intel Xeon Platinum 8180 processor (July 2017) : Tested by Intel as of July 11 th 2017: Platform: 2S Intel Xeon Platinum GHz (28 cores), HT disabled, turbo disabled, scaling governor set to performance via intel_pstate driver, 384GB DDR ECC RAM. CentOS Linux release (Core), Linux kernel el7.x86_64. SSD: Intel SSD DC S3700 Series (800GB, 2.5in SATA 6Gb/s, 25nm, MLC).Performance measured with: Environment variables: KMP_AFFINITY='granularity=fine, compact, OMP_NUM_THREADS=56, CPU Freq set with cpupower frequency-set -d 2.5G -u 3.8G -g performance. Caffe: ( revision f96b759f71b f690af267158b82b150b5c. Inference measured with caffe time --forward_only command, training measured with caffe time command. For ConvNet topologies, synthetic dataset was used. For other topologies, data was stored on local storage and cached in memory before training. Topology specs from (ResNet-50),and (ConvNet benchmarks; files were updated to use newer Caffe prototxt format but are functionally equivalent). Intel C++ compiler ver , Intel MKL small libraries version Caffe run with numactl -l. 5.7x inference throughput improvement on Intel Xeon Platinum 8180 processor (December 2018) with continued optimizations : Tested by Intel as of November 11 th 2018 :2 socket Intel(R) Xeon(R) Platinum GHz / 28 cores HT ON, Turbo ON Total Memory GB (12slots / 32 GB / 2666 MHz). CentOS Linux Core, kernel: el7.x86_64, SSD sda RS3WC080 HDD 744.1GB,sdb RS3WC080 HDD 1.5TB,sdc RS3WC080 HDD 5.5TB, Deep Learning Framework Intel Optimization for caffe version: 551a53d63a6183c233abaa1a19458a25b672ad41 Topology::ResNet_50_v1 BIOS:SE5C620.86B MKLDNN: 4e333787e0d66a1dca1218e99a891d493dbc8ef1 instances: 2 instances socket:2 (Results on Intel Xeon Scalable Processor were measured running multiple instances of the framework. Methodology described here: Synthetic data. Datatype: INT8 Batchsize=64 vs Tested by Intel as of July 11 th 2017:2S Intel Xeon Platinum GHz (28 cores), HT disabled, turbo disabled, scaling governor set to performance via intel_pstate driver, 384GB DDR ECC RAM. CentOS Linux release (Core), Linux kernel el7.x86_64. SSD: Intel SSD DC S3700 Series (800GB, 2.5in SATA 6Gb/s, 25nm, MLC).Performance measured with: Environment variables: KMP_AFFINITY='granularity=fine, compact, OMP_NUM_THREADS=56, CPU Freq set with cpupower frequency-set -d 2.5G -u 3.8G -g performance. Caffe: ( revision f96b759f71b f690af267158b82b150b5c. Inference measured with caffe time --forward_only command, training measured with caffe time command. For ConvNet topologies, synthetic dataset was used. For other topologies, data was stored on local storage and cached in memory before training. Topology specs from (ResNet-50). Intel C++ compiler ver , Intel MKL small libraries version Caffe run with numactl -l. 14x inference throughput improvement on Intel Xeon Platinum 8280 processor with Intel DL Boost: Tested by Intel as of 2/20/ socket Intel Xeon Platinum 8280 Processor, 28 cores HT On Turbo ON Total Memory 384 GB (12 slots/ 32GB/ 2933 MHz), BIOS: SE5C620.86B.0D (ucode: 0x200004d), Ubuntu LTS, kernel generic, SSD 1x sda INTEL SSDSC2BA80 SSD 745.2GB, nvme1n1 INTEL SSDPE2KX040T7 SSD 3.7TB, Deep Learning Framework: Intel Optimization for Caffe version: (commit hash: f159da247db3fe3a9d96a3116ca06b09a), ICC version , MKL DNN version: v0.17 (commit hash: 830a10059a018cd2634d cf2d8790a75a, model: BS=64, synthetic Data, 4 instance/2 socket, Datatype: INT8 vs Tested by Intel as of July 11 th 2017: 2S Intel Xeon Platinum GHz (28 cores), HT disabled, turbo disabled, scaling governor set to performance via intel_pstate driver, 384GB DDR ECC RAM. CentOS Linux release (Core), Linux kernel el7.x86_64. SSD: Intel SSD DC S3700 Series (800GB, 2.5in SATA 6Gb/s, 25nm, MLC).Performance measured with: Environment variables: KMP_AFFINITY='granularity=fine, compact, OMP_NUM_THREADS=56, CPU Freq set with cpupower frequency-set -d 2.5G -u 3.8G -g performance. Caffe: ( revision f96b759f71b f690af267158b82b150b5c. Inference measured with caffe time --forward_only command, training measured with caffe time command. For ConvNet topologies, synthetic dataset was used. For other topologies, data was stored on local storage and cached in memory before training. Topology specs from (ResNet-50),. Intel C++ compiler ver , Intel MKL small libraries version Caffe run with numactl -l. 2x More inference throughput improvement on Intel Xeon Platinum 9282 processor with Intel DL Boost : Tested by Intel as of 2/26/2019. Platform: Dragon rock 2 socket Intel Xeon Platinum 9282(56 cores per socket), HT ON, turbo ON, Total Memory 768 GB (24 slots/ 32 GB/ 2933 MHz), BIOS:SE5C620.86B.0D , Centos 7 Kernel el7.x86_64, Deep Learning Framework: Intel Optimization for Caffe version: d554cbf1, ICC , MKL DNN version: v0.17 (commit hash: 830a10059a018cd2634d cf2d8790a75a), model: BS=64, No datalayer syntheticdata:3x224x224, 56 instance/2 socket, Datatype: INT8 vs Tested by Intel as of July 11 th 2017: 2S Intel Xeon Platinum GHz (28 cores), HT disabled, turbo disabled, scaling governor set to performance via intel_pstate driver, 384GB DDR ECC RAM. CentOS Linux release (Core), Linux kernel el7.x86_64. SSD: Intel SSD DC S3700 Series (800GB, 2.5in SATA 6Gb/s, 25nm, MLC).Performance measured with: Environment variables: KMP_AFFINITY='granularity=fine, compact, OMP_NUM_THREADS=56, CPU Freq set with cpupower frequency-set -d 2.5G -u 3.8G -g performance. Caffe: ( revision f96b759f71b f690af267158b82b150b5c. Inference measured with caffe time --forward_only command, training measured with caffe time command. For ConvNet topologies, synthetic dataset was used. For other topologies, data was stored on local storage and cached in memory before training. Topology specs from (ResNet-50),. Intel C++ compiler ver , Intel MKL small libraries version Caffe run with numactl -l.

33 System Configuration: World Record + Real Workload Performance Leadership Performance results are based on testing as of dates shown in configuration and may not reflect all publicly available security updates. See configuration disclosure for details. No product or component can be absolutely secure. Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors. Performance tests, such as SYSmark and MobileMark, are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products. For more complete information visit 3.4X Facial Recognition for Microsoft: Tested by Intel as of 3/12/2019. Intel Xeon Platinum 8268 Processor, 24 cores, 384 GB (12 slots/ 32GB/ 2666 MHz), HT ON, BIOS: SE5C620.86B.BR , Ubuntu , , ngraph version: b dca9c63bf167e adf61995, ONNX version: 1.3.0, MKL DNN version: v0.18, MKLML_VERSION_ ,Topology: ResNet- 50,BS=1, Dataset: Synthetic, Datatype: INT8 w/ Intel DL Boost vs Tested by Intel as of 3/12/2019. Intel Xeon Platinum 8168 Processor, 24 cores, 384 GB (12 slots/ 32GB/ 2666 MHz), HT ON, BIOS: SE5C620.86B.BR , Ubuntu , , ngraph version: b dca9c63bf167e adf61995, ONNX version: 1.3.0, MKL DNN version: v0.18, MKLML_VERSION_ ,Topology: ResNet-50,BS=1, Dataset: Synthetic, Datatype: FP32 2.4x text detection performance improvement for JD.com: Tested by JD.com as of 1/27/ socket Intel Xeon Gold Processor, 24 cores HT On Turbo ON Total Memory 192 GB (12 slots/ 16GB/ 2666 MHz), CentOS el7.x86_64, Compiler: gcc 4.8.5, Deep Learning Framework: Intel Optimizations for Caffe with custom optimizations, Topology: EAST ( JD.com s private dataset, BS=1. Comparing performance on same system with FP32 vs INT8 w/ Intel DL Boost 2.01x medical image classification performance improvement for NeuSoft: Tested by Intel and HiSign as of 02/01/ socket Intel Xeon Platinum 8260 Processor, 24 cores HT On Turbo ON Total Memory 768 GB (12 slots/ 64GB/ 2666 MHz), BIOS version (ucode 0x400000A), RedHat 7.5 kernel el7.elrepo.x86_64, Compiler: gcc 4.8.5, Deep Learning Framework: Intel Optimizations for Caffe v1.1.2, Topology: modified Alexnet,custom dataset, BS=1. Comparing performance on same system with FP32 vs INT8 w/ Intel DL Boost 4.43X ML Inferencing for Target: Based on Intel Analysis on 2/16/ nd Gen Intel Xeon Platinum 8280 Processor (28 Cores) with 384GB, DDR4-2933, using Intel OpenVino 2019 R1. HT OFF, Turbo ON. CentOS Linux release , kernel el7.elrepo.x86_64. Topology: ResNet-50, dataset: Synthetic, BS=4 and 14 instance, Comparing FP32 vs Int8 w/ Intel DL Boost performance on the system. 3.26x latency reduction for Tencent* Cloud Video Analysis: Tested by Tencent as of 1/14/ socket Intel Xeon Gold Processor, 24 cores HT On Turbo ON Total Memory 192 GB (12 slots/ 16GB/ 2666 MHz), CentOS el7.x86_64, Compiler: gcc 4.8.5, Deep Learning Framework: Intel Optimizations for Caffe v1.1.3, Topology: modified inception v3, Tencent s private dataset, BS=1. Comparing performance on same system with FP32 vs INT8 w/ Intel DL Boost

34 System Configuration: SKUs Optimized for unique network needs Performance results are based on testing as of dates shown in configuration and may not reflect all publicly available security updates. See configuration disclosure for details. No product or component can be absolutely secure. Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors. Performance tests, such as SYSmark and MobileMark, are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products. For more complete information visit Up to 1.76x gains on networking workloads based on OVS DPDK: Tested by Intel on 1/21/ Node, 2x Intel Xeon Gold 6130 Processor on Neon City platform with 12x 16GB DDR4 2666MHz (384GB total memory), Storage: 1x Intel 240GB SSD, Network: 4x Intel XXV710-DA2, Bios: PLYXCRB1.86B.0568.D , ucode: 0x200004d (HT= ON, Turbo= OFF), OS: Ubuntu* with kernel: generic, Benchmark: Open Virtual Switch (on 4C/4P/8T 64B Mpacket/s), Workload version: OVS , DPDK , Compiler: gcc7.3.0, Other software: QEMU , VPP v18.10, Results: 9.6. Tested by Intel on 1/18/ Node, 2x Intel Xeon Gold 6230N Processor on Neon City platform with 12x 16GB DDR4 2999MHz (384GB total memory), Storage: 1x Intel 240GB SSD, Network: 6x Intel XXV710-DA2, Bios: PLYXCRB1.86B.0568.D , ucode: 0x (HT= ON, Turbo= OFF), OS: Ubuntu* with kernel: rc6-generic, Benchmark: Open Virtual Switch (on 6P/6C/12T 64B Mpacket/s), Workload version: OVS , DPDK , Compiler: gcc7.3.0, Other software: QEMU , VPP v18.10, Results: Tested by Intel on 1/18/ Node, 2x Intel Xeon Gold 6230N Processor with SST-BF enabled on Neon City platform with 12x 16GB DDR4 2999MHz (384GB total memory), Storage: 1x Intel 240GB SSD, Network: 6x Intel XXV710-DA2, Bios: PLYXCRB1.86B.0568.D , ucode: 0x (HT= ON, Turbo= ON (SST-BF)), OS: Ubuntu* with kernel: rc6-generic, Benchmark: Open Virtual Switch (on 6P/6C/12T 64B Mpacket/s), Workload version: OVS , DPDK , Compiler: gcc7.3.0, Other software: QEMU , VPP v18.10, Results: 16.9