amd epyc™ processors 卓越性能與應用...rom-56 - based on amd internal testing of altair...

1 | AMD EPYC™ 卓越性能與應用 @ VMWARE VFORUM | 2019 DECEMBER

[AMD Office Use Only - Approved for External Use]

1Confidential │ ©2019 VMware, Inc.

鄭凱元 Micky Cheng產品技術經理資料中心暨嵌入式解決方案事業群 AMD台灣

AMD EPYC™ Processors

卓越性能與應用


2 | AMD EPYC™卓越性能與應用@ VMWARE VFORUM | 2019 DECEMBER Roadmaps subject to change

P R O D U C T S



"ZEN2" IS A SUBSTANTIALLY HIGHER PERFORMING CORE WITH LOWER POWER CONSUMPTION VS. "ZEN"

7nm Process Technology Inverts the Competitive Playing Field and Enables Significant Compute Density

"ZEN2" Delivers Up to 4X Theoretical Peak FLOPS vs. "ZEN"

• 2X Cores Per SoC

• Improved Execution Pipeline

• 2X Integer

• 2X Load Store

• Higher Frequencies

• Up to 4X Shared L3 Cache (2X/Core)

2015 2016 2017 2018 2019

14 nm

14 Fin14+ Fin

10 Fin

12 nm

7 nm

Perf

orm

ance

/ W

att

Competitor Foundry

7 nm

(Same Performance)

~1.25x

See Endnote EPYC-07, ROM-04, ROM-37


4 | AMD EPYC™卓越性能與應用@ VMWARE VFORUM | 2019 DECEMBER

P R O D U C T S

*EPYC 7742 - see endnote ROM-114. **Not supported on all motherboards see ROM-06a. Also see endnote ROM-169.



✓

All-in feature set includes:

Up to 8 Channels of DDR4-3200*

4TB memory capacity

128 lanes PCIe® 4*

SMT & Turbo boost

18G Infinity Fabric™

Secure Memory Encryption

Secure Encrypted Virtualization

*A motherboard designed for 2nd Gen EPYC processors is required to enable all available functionality see endnote ROM-06a.



See endnote ROM-169, ROM-329



World records as of Nov 17, 2019. See AMD.com/worldrecords for details.

http://www.amd.com/worldrecords



6

9

12

World records as of Nov 17, 2019. See AMD.com/worldrecords for details.

http://www.amd.com/worldrecords



P R O D U C T S



▪

▪

▪

*A MOTHERBOARD DESIGNED FOR 2ND GEN EPYC PROCESSORS IS REQUIRED TO ENABLE ALL AVAILABLE FUNCTIONALITY See endnote ROM-06aSee endnote ROM-11, ROM-251



▪

▪

▪




®

See endnote ROM-42

AMD EPYC™ Performance on ANSYS Fluent® on YouTube

https://youtu.be/3d0RlJPFqJI



See endnote ROM-113

AMD EPYC™ Performance on GROMACS on YouTube

https://youtu.be/H6VAt5Fq9pI



See endnote ROM-112

AMD EPYC™ Performance on Blender on YouTube

https://youtu.be/7KJqzBjFbEE



See endnotes ROM-35, ROM-100

AMD EPYC™ Performance on VMmark® on YouTube

Read the AMD EPYC™ VMmark® 3 Performance Brief

https://youtu.be/voiOdVz8ojU

https://www.amd.com/system/files/documents/AMD-EPYC-New-World-Record-VMmark3.pdf



™


AMD EPYC™ Performance on VMmark® on YouTube

Read the AMD EPYC™ VMmark® 3 on vSAN® Performance Brief

https://youtu.be/voiOdVz8ojU

https://www.amd.com/system/files/documents/AMD-EPYC-Best-VMmark3-on-vSAN-Four-Nodes.pdf




AMD EPYC™ Performance on Docker® Containers on YouTube

Read the AMD EPYC™ Containers Performance Brief

https://youtu.be/J4gU0JgQSXk

https://www.amd.com/system/files/documents/EPYC-7002-Containers-Leadership-Performance.pdf



AMD EPYC™ Performance on Microsoft Azure D&E Series on YouTube

https://youtu.be/fiaYsqsQItU


20 | AMD EPYC™卓越性能與應用@ VMWARE VFORUM | 2019 DECEMBER•• See Endnote: Microsoft Azure HBv2

AMD EPYC™ Performance on Microsoft Azure HBv2 on YouTube

https://youtu.be/xdjYiERPRDE



AMD EPYC™ Performance on Virtual Desktops in Microsoft Azure® on YouTube

https://youtu.be/Pw9elIDYwpc


22 | AMD EPYC™卓越性能與應用@ VMWARE VFORUM | 2019 DECEMBER See Endnote: Amazon Web Services

AMD EPYC™ Performance on AWS on YouTube

https://youtu.be/UVNosSHp24Q



NVIDIA

USE OF THIRD-PARTY MARKS / LOGOS/ PRODUCTS IS FOR INFORMATIONAL PURPOSES ONLY AND NO ENDORSEMENT OF OR BY AMD IS INTENDED OR IMPLIED. GD-83



P R O D U C T S

BAIDU

Use of third-party marks/logos/products is for informational purposes only and no endorsement of or by AMD is intended or implied. GD-83



P R O D U C T S

Use of third-party marks/logos/products is for informational purposes only and no endorsement of or by AMD is intended or implied. GD-83



Intel® Xeon® Silver Processors

INTEL STACKS DO NOT INCLUDE XEON PLATINUM 9200 SERIES

SPEC®, SPEC CPU® and SPECrate® are registered trademarks of the Standard Performance Evaluation Corporation. More information about SPEC is available at www.spec.org

See endnote ROM-499, 500

http://www.spec.org/


27 | AMD EPYC™卓越性能與應用@ VMWARE VFORUM | 2019 DECEMBERSEE ENDNOTEs ROM-326, 327, 328


28 | AMD EPYC™卓越性能與應用@ VMWARE VFORUM | 2019 DECEMBERSEE ENDNOTE: ROM-329



Silver Processors

SPEC®, SPEC CPU® and SPECrate® are registered trademarks of the Standard Performance Evaluation Corporation. More information about SPEC is available at www.spec.org

SEE ENDNOTE: ROM-499

http://www.spec.org/



103

175

287 288

377

160

305

422

520

682

SEE ENDNOTES ROM-494, 495, 496, 497 AND 498

(est)

(est)



Altair® Radioss™ ANSYS® Fluent® Blender®

Microsoft SQL Server®Docker® Containers Apache Spark® PCIe® 4.0VMmark® 3.0

1.4X

GROMACS SPECvirt®

1.6XBEAMR

Azure® Cloud

See Endnotes ROM-36, ROM-42, ROM-56, ROM-99, ROM-112, ROM-113, ROM-131, ROM-136, ROM-173, ROM-175, ROM-266, ROM-271



EPYC-07 - Based on June 8, 2018 AMD internal testing of same-architecture product ported from 14 to 7 nm technology with similar implementation flow/methodology, using performance from SGEMM. EPYC-07

ROM-04 - Based on standard calculation method for determining FLOPS. ROM-04

ROM-06a - 2nd Gen AMD EPYC™ processors used on motherboards designed for the 1st Gen AMD EPYC processor require a BIOS update from your server manufacturer. The EPYC 7742, 7642 and 7542 are 225w parts and require additional updates, contact your server manufacturer for support. For PCIe® 4 and DDR4-3200 memory support, please contact your server manufacturer. A motherboard designed for 2nd Gen EPYC processors is required to enable all available functionality. ROM-06a

ROM-11 - EPYC™ 7002 series has 8 memory channels, supporting 3200 MHz DIMMs yielding 204.8 GB/s of bandwidth vs. the same class of Intel Scalable Gen 2 processors with only 6 memory channels and supporting 2933 MHz DIMMs yielding 140.8 GB/s of bandwidth. 204.8 / 140.8 = 1.454545 - 1.0 = .45 or 45% more. AMD EPYC has 45% more bandwidth. Class based on industry-standard pin-based (LGA) X86 processors. ROM-11

ROM-35 - AMD 7702 score of 12.88@14 tiles on 2 node 4 socket VMmark 3.1, classic SAN. can be found at https://www.vmware.com/content/dam/digitalmarketing/vmware/en/pdf/vmmark/2019-08-07-HPE-ProLiant-DL385Gen10.pdf. Product available Aug 7, 2019. The next highest score, 9.02@9 tiles, with the 8280, can be found at https://www.vmware.com/products/vmmark/results3x.0.html VMware VMmark 3.x results can be found at https://www.vmware.com/products/vmmark/results3x.html. ROM-35

ROM-36 - The AMD EPYC 7702 server is the only 4 host 4 socket VMmark 3.1 vSAN score of 12.23 @ 13 tiles https://www.vmware.com/content/dam/digitalmarketing/vmware/en/pdf/vmmark/2019-08-07-HPE-ProLiant-DL325Gen10.pdf. The current #1 score for 4 node 8 sockets or less is 8.51 @ 9 tiles with the Gold 6152, https://www.vmware.com/products/vmmark/results3x.0.html. AMD 7702 score of can be found at https://www.vmware.com/content/dam/digitalmarketing/vmware/en/pdf/vmmark/2019-08-07-HPE-ProLiant-DL325Gen10.pdf. ROM-36

ROM-37 - Results as of 8/7/2019 using SPECrate(R)2017_int_base. The EPYC 7742 2P score is 682 on the SPECrate®2017_int_base, https://spec.org/cpu2017/results/res2019q3/cpu2017-20190722-16242.html. EPYC 7601 2P score of 304 results at http://spec.org/cpu2017/results/res2019q2/cpu2017-20190411-11817.pdf. 682 / 304 = 2.24 or 2.2x the integer performance for the EPYC 7742. SPEC®, SPECrate® and SPEC CPU® are registered trademarks of the Standard Performance Evaluation Corporation. See www.spec.org for more information. ROM-37

ROM-42 - Based on AMD internal testing of ANSYS FLUENT 19.1 as of July 17, 2019. 885 rating on lm6000_16m benchmark using 2 x AMD EPYC™ 7742 processors in AMD "Daytona" Reference Design, 1TB (16 x 64GB DDR4-3200) memory, 256GB SATA (OS), 1TB NVMe (Data), Red Hat® Enterprise Linux v7.6, ANSYS FLUENT 19.1, NPS = NPS4, SMT = OFF, BOOST = ON, X2APIC = ON, Determinism Slider = Performance, Governor = Performance, CC6 = Disabled. 444 rating on lm6000_16m benchmark using 2 x Intel Xeon Platinum 8280 processors in Supermicro SuperServer 2029U-TN24R4T server, 384GB (12 x 32GB DDR4-2933) memory, 240GB SATA (OS), 500GB NVMe (Data), Red Hat® Enterprise Linux v7.6, ANSYS FLUENT 19.1, Power Management = Max Performance, Intel Hyper-Threading Technology = OFF, SNC = ON, ADDDC = OFF. Testing with other AMD EPYC or Intel Xeon processors may result in different performance results. ROM-42

ROM-56 - Based on AMD internal testing of Altair RADIOSS 2018 as of July 17, 2019. 9.75 rating on T10M benchmark using 2 x AMD EPYC™ 7742 processors in AMD "Daytona" Reference Design, 1TB (16 x 64GB DDR4-3200) memory, 256GB SATA (OS), 1TB NVMe (Data), Red Hat® Enterprise Linux v7.6, Altair RADIOSS 2018, NPS = NPS4, SMT = OFF, BOOST = ON, X2APIC = ON, Determinism Slider = Performance, Governor = Performance, CC6 = Disabled. 5.68 rating on T10M benchmark using 2 x Intel Xeon Platinum 8280 processors in Supermicro SuperServer 2029U-TN24R4T server, 384GB (12 x 32GB DDR4-2933) memory, 240GB SATA (OS), 500GB NVMe (Data), Red Hat® Enterprise Linux v7.6, Altair RADIOSS 2018, Power Management = Max Performance, Intel Hyper-Threading Technology = OFF, SNC = ON, ADDDC = OFF. Testing with other AMD EPYC or Intel Xeon processors may result in different performance results. ROM-56

ROM-99 - A 2P EPYC 7702 powered server has SPECvirt_sc2013 score of 5451.2 and 305 VMs, https://www.spec.org/virt_sc2013/results/res2019q3/virt_sc2013-20190726-00122-perf.html as of August 7, 2019. The next highest score is a 2P Intel Platinum 8180 server with a score of 3376 and 189 VMs, https://www.spec.org/virt_sc2013/results/res2017q4/virt_sc2013-20171017-00098-perf.html as of July 28, 2019. SPEC® and SPECvirt® are registered trademarks of the Standard Performance Evaluation Corporation. See www.spec.org for more information. ROM-99

ROM-100 - Based on VMmark 3.1 SAN. An EPYC 7702 powered server delivered a VMmark 3.1 SAN storage score of 12.88 with 14 tiles on Aug 7, 2019. https://www.vmware.com/content/dam/digitalmarketing/vmware/en/pdf/vmmark/2019-08-07-HPE-ProLiant-DL385Gen10.pdf. Product available Aug 7, 2019. The next highest 2 node 4 socket score of 9.02 @ 9 tiles with Platinum 8280, ttps://www.vmware.com/products/vmmark/results3x.0.html, July 28, 2019. VMware VMmark 3.0 and 3.1 results can be found at https://www.vmware.com/products/vmmark/results3x.html. ROM-100

ROM-112 - Based on AMD internal testing as of 30July2019 of Blender version 2.8RC3 on Ubuntu 19.04 – kernel 5.0x, rendering a frame from the Blender Open Source Movie Project “Spring” Shot file 02_040_A with 3000 samples, denoising disabled, motion blur disabled and render time added to EXIF data on AMD reference platform configured with 2 x AMD EPYC 7742 processors, 16 x 32GB, 3200MHz DDR4 DIMMs, 128GB drive, AMD Radeon RX 580 graphics card and AGESA 1.0.0.1 BIOS versus an Intel server from a major OEM configured with 2 x Intel Platinum 8280 processors, 12 x 32GB, DDR4 2933MHz DIMMs, 256GB drive, AMD Radeon RX 580 graphics card and 2.2.10 BIOS. The AMD system completed the frame render in an average of 24.90 mins and the Intel system completed the same render in an average of 47.58 mins. 47.58/24.90=1.91-1=91% faster or 1.9X the performance. Results may vary. ROM-112

ROM-113 - Based on AMD internal testing as of 30July2019 of GROMACS 2019.3 lignocellulose3M_rf benchmark. AMD reference platform configured with 2 x AMD EPYC 7742 processors, 16 x 64GB, 3200MHz DDR4 DIMMs, 256GB OS drive, 1TB NVMe data drive and AGESA 1.0.0.1 BIOS versus an Intel server from a major OEM configured with 2 x Intel Platinum 8280 processors, 12 x 64GB, DDR4 2933MHz DIMMs, 240GB OS drive and a 500GB NVMe drive. The AMD platform achieved 5.3 ns/day and the Intel platform achieved 3.3 ns/day of simulation time. 5.3/3.3=1.6-1=60% better or 1.6x the performance. Results may vary. ROM-113



ROM-114 - An 2P EPYC 7742 powered server has SPECrate®2017_int_peak score of 749 and an int_base score of 682, URL. The next highest int_peak score is a 2P Intel Platinum 9282 server with a score of 676 and an int_base score of 643, http://spec.org/cpu2017/results/res2019q3/cpu2017-20190624-15369.pdf. An 1P EPYC 7742 powered server has SPECrate®2017_int_peak score of 385 and an int_base score of 349, URL. The next highest int_peak score is a 1P Intel Platinum 8280L server with a score of 180, http://spec.org/cpu2017/results/res2019q2/cpu2017-20190319-11289.pdf. The next highest int_bK113:K115ase score is a 1P Intel Platinum 8280 server with a score of 181, http://spec.org/cpu2017/results/res2019q2/cpu2017-20190318-11230.pdf. All Intel scores as of July 28, 2019. SPEC®, SPECrate® and SPEC CPU® are registered trademarks of the Standard Performance Evaluation Corporation. See www.spec.org for more information. ROM-114

ROM-115 - 84% higher JBB MultiJVM max – Footnote: An EPYC™ 7702P powered server has a World Record 1P SPECjbb2015-MultiJVM Max score of 171,634 (SPECjbb2015-MultiJVM Critical scored 75,275), http://www.spec.org/jbb2015/results/res2019q3/jbb2015-20190717-00467.html. 84% higher than the previous #1, a server with an Intel Platinum 8280 with a score of 93,484 (SPECjbb2015-MultiJVM Critical scored 25,413), https://www.spec.org/jbb2015/results/res2019q2/jbb2015-20190313-00372.html. SPEC® and SPECjbb® are registered trademarks of the Standard Performance Evaluation Corporation. See www.spec.org for more information. ROM-115

ROM-131 - Based on AMD internal testing as of 30July2019 of HSsort test from TPCx-HS benchmark version 2.0.3 on 2 Terabytes of data. 3 x Supermicro AS -1123US-TN10RT servers configured with 2 x AMD EPYC 7542 processors 32c per socket/64c per node16x 32GB 3200MHz DDR4 DIMMs, 3x 1TB NVMe drives, Mellanox CX-5 NIC and BIOS 2.0 NPS= NPS4; SMT=ON; Determinism Slider= Power; versus an 3 x Intel servers from a major OEM configured with 2 x Intel Platinum 8280 processors, 28c per socket/56c per node, 12x 32GB DDR4 2933MHz DIMMs, 3x 1TB NVMe drives, Mellanox CX-5 NIC and BIOS 3.1 Hyper-Threading=Enable; Power Technology: Custom; Power Performance Tuning: Bios Controls EPB; ENERGY_PERF_BIAS_CFG mode: Extreme Performance. Both systems used RHEL 7.6, openjdkversion1.8.0_212-b04, HDP 3.1, Spark version 2.3.2. The AMD servers achieved an average of 2,301,033,506 b/s and the Intel servers achieved an average of 1,500,587,408 b/s. 2,301,033,506/1,500,587,408 = 1.53-1=53% more throughput or 1.53x the throughput. Results may vary. ROM-131

ROM-136 - Based on AMD internal testing as of 02Aug2019 of HammerDB Benchmark v3.1 on Microsoft SQL Server®. AMD reference platform with 2 x AMD EPYC 7742 processors, 16 x 32GB, 3200MHz DDR4 DIMMs, 6 x Micron 9300 2.9TB NVMedrives and BIOS AGESA 1.0.0.1 versus an Intel server from a major OEM configured with 2 x Intel Platinum 8280 processors, 4x64GB and 8x32GB 2933MHz DDR4 DIMMs, 6 x Micron 9300 2.9TB NVMe drives and BIOS 3.1. The AMD server achieved an average of 8,935,583 TPM and the Intel server achieved an average of 6,132,927 TPM. 8,935,583/6,132,927 = 1.46-1=46% more throughput/TPM or 1.5x the throughput/TPM. Results may vary. ROM-136

ROM-138 - Results as of 8/7/2019 using VMmark(R) 3.1 vSAN. AMD 7702 score of can be found at https://www.vmware.com/content/dam/digitalmarketing/vmware/en/pdf/vmmark/2019-08-07-HPE-ProLiant-DL325Gen10.pdf. Product available Aug 7, 2019. The next highest score with an Intel 6152, can be found at, ttps://www.vmware.com/products/vmmark/results3x.0.html. VMware VMmark 3.x results can be found at https://www.vmware.com/products/vmmark/results3x.html. ROM-138

ROM-139 - World record database performance – Footnote: Results as of 8/7/2019. 8-node EPYC™ 7702 result published at TPC website http://www.tpc.org/3341. Previous #1 published result on the TPC website at: http://www.tpc.org/3306. Product availability 8/7/2019. TPC and TPC Benchmark are registered trademarks of the Transaction Processing Performance Council. ROM-139

ROM-169 - For a complete list of world records see http://amd.com/worldrecords. ROM-169

ROM-172 - Compares running 8,400 identical Docker containers on Intel 2 Socket Xeon 8280 based systems versus AMD EPYC™ 7702 based. Internal AMD testing performed on each system running from 4 to 256 containers with identical loads, relative performance/throughput in events per second. The Intel based system reached 60% and 100% thread saturation at between 64 to 88 containers and 96 to 128 containers respectively. The AMD based system reached 60% thread saturation at between 140 to 185 containers. 100% of threads on the AMD system were saturated at 256 containers. At the closest measured point to 60% thread saturation, the Intel based system ran 64 containers while the AMD EPYC™ based system ran 140 containers within +/- 3% performance calculated as events processed per second per container. This resulted in the AMD EPYC™ based systems running 2.2X (140 vs 64) the number of containers at nominal (60%) utilization. System Configurations:Intel Xeon based servers with (2) Intel® Xeon® Platinum 8280 2.7G, 28C/56T, (12) 32GB RDIMM, 2933MT/s, Dual Rank DIMMs, Hard Drives including (1) 480GB SSD SAS Mix Use, (1) 1.6TB NVME Drive; AMD EPYC™ servers with (2) AMD EPYC™ 7702, 2.0G, 64C/128T, (16) 32GB RDIMM, 2933MT/s, Dual Rank DIMMs, Hard Drives including (1) 480GB SSD SAS Mix Use, (1) 1.6TB NVME Drive. ROM-172

ROM-173 - AMD internal testing completed on 29Jul2019 on AMD reference platform configured with 2 x EPYC 7742 processors, 16 x 32GB DDR4 2933MHz memory, 6 x Samsung PCIe Gen4 PM1733 NVMe 3.84TB drives, Ubuntu 19.04, 5.0.0-20-generic kernel compared to an Intel server from a major OEM configured with 2 x Intel Platinum 8280 processors, 12 x DDR4 2933 memory, 6 x Samsung PM1725b 1.6 TB (U.2) - Gen3 drives, Ubuntu 19.04, 5.0.0-20-generic kernel. Results measured via the fio-3.12 benchmark. AMD achieved 43.1 GB/s on Sequential Reads using 128K I/O size and the Intel system achieved 20.9 GB/s on Sequential Reads using 128K I/O size. 43.1/20.9=2.1X. ROM-173

ROM-175 - Testing completed on 29Jul2019 on AMD reference platform configured with 2 x EPYC 7742 processors, 16 x 32GB DDR4 2933MHz memory, 6 x Samsung PCIe Gen4 PM1733 NVMe 3.84TB drives, Ubuntu 19.04, 5.0.0-20-generic kernel compared to an Intel server from a major OEM configured with 2 x Intel Platinum 8280 processors, 12 x DDR4 2933 memory, 6 x Samsung PM1725b 1.6 TB (U.2) - Gen3 drives, Ubuntu 19.04, 5.0.0-20-generic kernel. Results measured via the fio-3.12 benchmark. AMD achieved 21.5 GB/s on Sequential writes using 128K I/O size and the Intel system achieved 11.5 GB/s on Sequential writes using 128K I/O size. 21.5/11.5=1.9X. ROM-175

ROM-250 - Results as of 8/7/2019 using SPECrate(R)2017_int_base. EPYC 7742 score of 682, https://spec.org/cpu2017/results/res2019q3/cpu2017-20190722-16242.html. Intel Platinum 8280L score 364, http://spec.org/cpu2017/results/res2019q2/cpu2017-20190429-12779.pdf, Aug 2, 2019. SPEC®, SPECrate® and SPEC CPU® are registered trademarks of the Standard Performance Evaluation Corporation. See www.spec.org for more information. ROM-250

ROM-251 - Results as of 8/7/2019 using SPECrate(R)2017_fp_base. EPYC 7742 score of 524, https://spec.org/cpu2017/results/res2019q3/cpu2017-20190722-16241.html. Intel Platinum 8280L score 301, https://spec.org/cpu2017/results/res2019q2/cpu2017-20190429-12775.pdf, Apr, 2019. SPEC®, SPECrate® and SPEC CPU® are registered trademarks of the Standard Performance Evaluation Corporation. See www.spec.org for more information. ROM-251



ROM-266 - Compares running 8,400 identical Docker containers on Intel 2 Socket Xeon 8280 based systems versus AMD EPYC™ 7702 based. Internal AMD testing performed on each system running from 4 to 256 containers with identical loads, relative performance/throughput in events per second. The Intel based system ran 256 containers at 1595 events/sec while the AMD EPYC™ based system ran 256 containers at 2912 events/sec. The AMD system achieved ~83% better performance calculated as events processed per second per container. (2912-1595)/1595 = 83%. System Configurations: Intel Xeon based servers with (2) Intel® Xeon® Platinum 8280 2.7G, 28C/56T, (12) 32GB RDIMM, 2933MT/s, Dual Rank DIMMs, Hard Drives including (1) 480GB SSD SAS Mix Use, (1) 1.6TB NVME Drive; AMD EPYC™ servers with (2) AMD EPYC™ 7702, 2.0G, 64C/128T, (16) 32GB RDIMM, 2933MT/s, Dual Rank DIMMs, Hard Drives including (1) 480GB SSD SAS Mix Use, (1) 1.6TB NVME Drive. Both systems were configured with Red Hat® Enterprise Linux® 8, kernel 4.18.0-80.el8.x86_64 and Docker Engine 18.06.3-ce and default BIOS settings. ROM-266

ROM-271 - AMD internal testing completed on 06Aug2019 on AMD reference platform configured with 2 x EPYC 7742 and a Mellanox ConnectX-6 InfiniBand using Windows 2019 compared to an Intel server from a major OEM configured with 2 x Intel Platinum 8280 processors and a Mellanox ConnnectX-6 using Windows 2019. Results are an overage of bi-directional bandwidth measured with InfiniBand read bandwidth at 32K block size. The AMD platform measured 395Gb/s and the Intel system measured 202 Gb/s. 395/202=2X ROM-271

ROM-326 - Hardware requirements: To deliver 3200 VMs (virtual machines) with 1 core and 8GB DDR4 DRAM memory requires 50 single socket AMD EPYC 7702P CPU servers with 512 GB of memory and for Intel 100 dual socket Gold 6242 (16 cores ea.) based servers. 3200 / 64 cores per AMD server = 50 AMD servers. 3200 / 32 cores per Intel server = 100 Intel servers. Power estimates: AMD 315 watts per server per hour, for a total solution power of 1360.8 kW per month for 50 servers. Intel 415 watts per server per hour, for a total solution power of 3585.6 kW per month for 100 servers. Three year total power with a PUE of 2 and a power cost of $0.12 per kWh: AMD - $97,977.60 and Intel - $258,163.20. $97,977.60 / $258,163.20 = 38% or AMD has 62% lower power cost. Over 3 years AMD power consumption is 408,240 kWh; Intel consumption is 1,075,680 kWh. 408,240 / 1,075,680 = 38% or AMD has an estimated 62% lower power consumption. Disclaimer: This scenario contains many assumptions and estimates and, while based on AMD internal research and best approximations, should be considered an example for information purposes only, and not used as a basis for decision making in place of actual testing. ROM-326

ROM-327 - Example to deliver 3200 VMs (virtual machines) with virtualization software licensed by the socket, and providing 1 core and 8GB DDR4 DRAM memory per VM, would require 50 single socket AMD EPYC 7702P CPU servers with 512 GB of memory (3200/64 cores ea. = 50); and 100 Intel dual socket Gold 6242 (16 cores ea.,3200 / 32 cores = 100). 50 one socket servers = 50 AMD sockets. 100 two socket servers = 200 Intel sockets. For socket-based licensing: 50 / 200 = 25%, resulting in 1P AMD EPYC 7702P servers using 75% fewer virtualization software per-socket licenses than the 2P Intel 6242 servers for 3200VMs. ROM-327

ROM-328 - To deliver 3200 VMs (virtual machines) with 1 core and 8GB DDR4 DRAM memory requires 50 single socket AMD EPYC 7702P CPU servers with 512 GB of memory and for Intel 100 dual socket Gold 6242 (16 cores ea.) based servers. 3200 / 64 cores per AMD server = 50 AMD servers. 3200 / 32 cores per Intel server = 100 Intel servers. 50 / 100 = 50% fewer server needed with AMD. ROM-328

ROM-329 - Comparison results calculated by AMD based on a hypothetical scenario hosting 3,200 virtual machines on 2-socket Intel Xeon 6242 16 cores ea. (32 total cores) based systems compared to 1-socket AMD EPYC™ 7702P 64 total core

system, each VM (virtual machine) allocated one core, (1) 1 TB hard drive @ $387, and 8GB of DRAM with a cost of $16.41 per GB using 16GB DDR4 DRAM. All calculations are based on AMD's best estimates of what actual costs and other values

will be for both AMD and Intel based platforms. Each AMD 7702P system supports 64 VMs and each Intel 6242 system supports 32 VMs, for a total of 50 AMD powered servers and 100 Intel based servers. These estimates do not include and VM

server management costs. System Configurations and Hardware Costs: Intel Xeon based servers include (2) Intel® Xeon® Gold 6242 @ $2,529 ea,($5,058) with 256 GB of memory, a dual socket 2U rack mount server, each with power supplies, and

NIC @ $2,500, for a total per server price of $12,146.56 ea. and a total hardware acquisition price of $1,214,656 for 100 servers. AMD EPYC™ servers include a single socket 2U rack mount chassis, with (1) AMD EPYC™ 7702P at $4,425, with 512

GB of memory, a single socket 2U rack mount server, each with power supplies, and NIC @ $2,200, for a total per server price of $15,414.52 ea. and a total hardware acquisition price of $770,726 for 50 servers. Power estimates: AMD 315 watts per

server per hour, for a total solution power of 1360.8 kW per month for 50 servers. Intel 415 watts per server per hour, for a total so-lution power of 3585.6 kW per month for 100 servers. Three year total power with a PUE of 2 and a power cost of $0.12

per kWh: AMD - $97,977.60 and Intel - $258,163.20. Real estate Data center 3 year real estate cost estimates based on $20/mth/sq ft and 27 sq ft per rack are for: Intel is 5 rack cabinets at $97,200 and for AMD of $48,600 (2.5 rack cabinets). Server

Administration cost is calculated with an estimate of $110,500 annually per server administrator (includes 30% burden) with a ratio of one server administrator per 30 servers resulting in Intel cost of $1,105,000 (for 100 servers) and for AMD $552,500

(for 50 servers) for total 3 yr server administration costs. Total estimated 3 Year TCO as a result is $2,675,019 for the Intel-based Systems and $1,469,804 for the AMD EPYC powered systems. As a result, AMD EPYC based systems are estimated to

deliver up to a 45% lower TCO (excluding software costs). Annual cost per VM: (1469804/3200)/3=$279; (2675091/3200)/3=$153. Disclaimer: This scenario contains many assumptions and estimates and, while based on AMD internal research and

best approximations, should be considered an example for information purposes only, and not used as a basis for decision making in place of actual testing. ROM-329

ROM-494

Projections as of August 2, 2019 an AMD EPYC 7272 powered server will have a estimated SPECrate®2017_int_base score of 160 and cost $625 per processor. An Intel Silver 4215 based server has a score of 103

(http://spec.org/cpu2017/results/res2019q3/cpu2017-20190805-16603.pdf), and costs $794 per processor as of Oct 17, 2019. Intel ppd 103 / [$794*2] = 0.065. AMD ppd 160 / [$625*2] = 0.13. 0.13 / 0.065 = 2.0X (200%) the estimated

performance/dollar with AMD. AMD 1kU prices as of Dec 4, 2019; Intel prices as of Dec 4, 2019 from https://ark.intel.com/. All prices USD. SPEC score as of Oct 17, 2019. SPEC®, SPECrate® and SPEC CPU® are registered trademarks of the

Standard Performance Evaluation Corporation. See www.spec.org for more information. ROM-494

ROM-495

An AMD EPYC 7352 powered server has a SPECrate®2017_int_base score of 305 (http://spec.org/cpu2017/results/res2019q4/cpu2017-20190930-18854.pdf) and cost $1350 per processor. An Intel Gold 6226 based server has a score of 175

(http://spec.org/cpu2017/results/res2019q3/cpu2017-20190709-16044.pdf), and cost $1776 per processor as of Oct 17, 2019. Intel ppd 175 / [$1776*2)]= 0.05. AMD ppd 305 / [$1350*2] = 0.11. 0.11 / 0.05 = 2.3X (230%) the performance/dollar with

AMD. AMD 1kU prices as of Dec 4, 2019; Intel prices as of Dec 4, 2019 from https://ark.intel.com/. All prices USD. SPEC scores as of Oct 17, 2019. SPEC®, SPECrate® and SPEC CPU® are registered trademarks of the Standard Performance

Evaluation Corporation. See www.spec.org for more information. ROM-495



ROM-496

An AMD EPYC 7502 powered server has a SPECrate®2017_int_base score of 422 (http://spec.org/cpu2017/results/res2019q4/cpu2017-20190916-17922.pdf) and cost $2600 per processor. An Intel Gold 6252 based server has a score of 287

(http://spec.org/cpu2017/results/res2019q4/cpu2017-20190916-18249.pdf), and cost $3655 per processor as of Oct 17, 2019. Intel ppd 287 / [$3655*2] = 0.04. AMD ppd 422 / [$2600*2] = 0.08. 0.08 / 0.04 = 2.0X (200%) the performance/dollar with

AMD. AMD 1kU prices as of Dec 4, 2019; Intel prices as of Dec 4, 2019 from https://ark.intel.com/. All prices USD. SPEC scores as of Oct 17, 2019. SPEC®, SPECrate® and SPEC CPU® are registered trademarks of the Standard Performance


ROM-497

Projections as of August 2, 2019 an AMD EPYC 7642 powered server will have a estimated SPECrate®2017_int_base score of 520 and cost $4775 per processor. An Intel Platinum 8260 based server has a score of 288

(http://spec.org/cpu2017/results/res2019q3/cpu2017-20190703-15911.pdf), and cost $4702 per processor as of Oct 17, 2019. Intel ppd 288 / [$4702*2] = 0.03. AMD ppd 520 / [$4775*2] = 0.054. 0.054 / 0.03 = 1.8X (180%) the estimated

performance/dollar with AMD. AMD 1kU prices as of Dec 4, 2019; Intel prices as of Dec 4, 2019 from https://ark.intel.com/. SPEC score as of Oct 17, 2019. All prices USD. SPEC®, SPECrate® and SPEC CPU® are registered trademarks of the


ROM-498

An AMD EPYC 7742 powered server has a SPECrate®2017_int_base score of 682 (http://spec.org/cpu2017/results/res2019q3/cpu2017-20190722-16242.pdf) and cost $6950 per processor. An Intel Platinum 8280M based server has a score of 377

(http://spec.org/cpu2017/results/res2019q3/cpu2017-20190709-16044.pdf), and cost $13,012 per processor as of Oct 17, 2019. Intel ppd 377 / [$13,012*2] = 0.014. AMD ppd 682 / [$6950*2] = 0.05. 0.05 / 0.014 = 3.4X (340%) the performance/dollar

with AMD. AMD 1kU prices as of Dec 4, 2019; Intel prices as of Dec 4, 2019 from https://ark.intel.com/. All prices USD. SPEC scores as of Oct 17, 2019. SPEC®, SPECrate® and SPEC CPU® are registered trademarks of the Standard Performance


ROM-499

Both published and estimated SPECrate®2017_int_base performance as of Oct 17, 2019. AMD EPYC 2P processor scores reflect highest scoring processor per core count: 7742 (64c) score of 682 (http://spec.org/cpu2017/results/res2019q3/cpu2017-

20190722-16242.pdf); 7642 (48c) est. score of 520; 7542 (32c) score of 436 (http://spec.org/cpu2017/results/res2019q3/cpu2017-20190902-17339.pdf); 7302 (16c) score of 244 (http://spec.org/cpu2017/results/res2019q3/cpu2017-20190902-

17342.pdf); 7262 (8c) score of 134 (http://spec.org/cpu2017/results/res2019q3/cpu2017-20190902-17341.pdf). Intel® Xeon® Scalable 2P processors scores using the highest score in each metal class and series excluding the 9200 series: Platinum

8280L score of 364 (http://spec.org/cpu2017/results/res2019q2/cpu2017-20190429-12779.pdf); Gold 6252 score of 287 (http://spec.org/cpu2017/results/res2019q4/cpu2017-20190916-18249.pdf); Gold 5220 score of 212

(http://spec.org/cpu2017/results/res2019q3/cpu2017-20190709-16048.pdf); Silver 4216 score of 186 (http://spec.org/cpu2017/results/res2019q3/cpu2017-20190805-16606.pdf); Bronze 3106 score of 45.5

(http://spec.org/cpu2017/results/res2018q1/cpu2017-20171225-02062.pdf). Estimate as of Aug 4, 2019 for AMD EPYC 7642 (48C) processors using preproduction parts and SPECrate®2017_int_base internal testing results. Results may vary with

production silicon testing. SPEC®, SPECrate® and SPEC CPU® are registered trademarks of the Standard Performance Evaluation Corporation. See www.spec.org for more information. ROM-499

ROM-500

Published SPECrate®2017_int_base performance as of Oct 17, 2019. AMD EPYC 1P processor scores reflect highest scoring processor per core count: 7742 (64c) score of 353 (http://spec.org/cpu2017/results/res2019q3/cpu2017-20190830-

17269.pdf); 7542 (32c) score of 218 (http://spec.org/cpu2017/results/res2019q3/cpu2017-20190902-17505.pdf); 7302P (16c) score of 122 (http://spec.org/cpu2017/results/res2019q3/cpu2017-20190902-17494.pdf); 7262 (8c) score of 67.1

(http://spec.org/cpu2017/results/res2019q3/cpu2017-20190902-17486.pdf). Intel® Xeon® Scalable 1P processors scores using the highest score in each metal class and series excluding the 9200 series: Platinum 8280 score of 181

(http://spec.org/cpu2017/results/res2019q2/cpu2017-20190318-11230.pdf); Gold 6212U score of 143 (http://spec.org/cpu2017/results/res2019q2/cpu2017-20190527-14835.pdf); Silver 4216 score of 87.3

(http://spec.org/cpu2017/results/res2019q2/cpu2017-20190524-14372.pdf); Bronze 3106 score of 22.7 (http://spec.org/cpu2017/results/res2018q3/cpu2017-20180809-08226.pdf). SPEC®, SPECrate® and SPEC CPU® are registered trademarks of the


Microsoft Azure HBv2

Over 50% more memory bandwidth based on Microsoft internal testing on a Microsoft Azure server platform which measured 356 GB/s on STREAM Triad compared to Intel Platinum 8260 platform with measured ~220 GB/s based on testing by

independent 3rd party reviewers. 4 teraFLOPS per VM based Microsoft internal testing using Microsoft Azure server platform with an HBv2 VM with 120 cores from 2 x EPYC 7002 series CPUs.

Amazon Web Services

The newest EC2 instances are powered by custom AMD EPYC processors running at 2.5 GHz and are priced 10% lower than comparable instances. https://aws.amazon.com/blogs/aws/new-lower-cost-amd-powered-ec2-instances/;

https://aws.amazon.com/ec2/pricing/on-demand/; The information provided here is for information purposes only. AMD DISCLAIMS ALL WARRANTIES, implied warranties or guarantees of any kind, including but not limited to the suitability or fitness of

any product mentioned here for any purpose. This demonstration tool is intended to illustrate the comparative differences between the products shown. Prices are sourced in US$ from https://aws.amazon.com/ec2/pricing/on-demand/. As advertised,

Amazon EC2 offers 10% lower cost compute and memory with AMD processors compared to comparable instances.. Actual prices and performance will vary at any given time. The tool does not constitute an offer to buy nor sell any product shown. IN

NO EVENT SHALL AMD BE LIABLE FOR ANY DAMAGES, WHETHER THOSE DAMAGES ARE DIRECT, CONSEQUENTIAL, INCIDENTAL, OR SPECIAL, FLOWING FROM YOUR USE OF OR INABILITY TO USE THE TOOL OR INFORMATION

PROVIDED HEREWITH OR RESULTS OF THE TOOL'S USE EVEN IF AMD HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES.



DISCLAIMER

The information contained herein is for informational purposes only, and is subject to change without notice. While every precaution has been taken in the preparation of this document, it may contain technical inaccuracies, omissions and typographical errors, and AMD is under no obligation to update or otherwise correct this information. Advanced Micro Devices, Inc. makes no representations or warranties with respect to the accuracy or completeness of the contents of this document, and assumes no liability of any kind, including the implied warranties of noninfringement, merchantability or fitness for particular purposes, with respect to the operation or use of AMD hardware, software or other products described herein. No license, including implied or arising by estoppel, to any intellectual property rights is granted by this document. Terms and limitations applicable to the purchase or use of AMD’s products are as set forth in a signed agreement between the parties or in AMD's Standard Terms and Conditions of Sale. GD-18

©2019 Advanced Micro Devices, Inc. All rights reserved. AMD, the AMD Arrow logo, EPYC, the EPYC logo, Radeon, Radeon Instinct and combinations thereof are trademarks of Advanced Micro Devices, Inc. Altair and Radioss Radioss are trademarks or registered trademarks of Altair Engineering, Inc. ANSYS, FLUENT and any and all ANSYS, Inc. brand, product, service and feature names, logos and slogans are registered trademarks or trademarks of ANSYS, Inc. or its subsidiaries in the United States or other countries. Blender is a registered trademark of the Blender Foundation in EU and USA. Docker and the Docker logo are trademarks or registered trademarks of Docker, Inc. Apache Spark is a trademark of The Apache Software Foundation. Microsoft, Azure and SQL Server are registered trademarks of Microsoft in the US and/or other countries., Vmmark, VMware and vSAN are trademarks or registered trademarks of VMware in the US or other countries. PCIe is a registered trademark of PCI-SIG Corporation. Other product names used in this publication are for identification purposes only and may be trademarks of their respective companies.

amd epyc™ processors 卓越性能與應用...rom-56 - based on amd internal testing of altair...

Documents