Data Center and AI Premiere Endnotes

SP5-011E: SPECpower_ssj®2008 comparison based on published 2P server results as of 6/13/2023. Configurations: 2P AMD EPYC 9654 (30,602 overall ssj_ops/W, 2U, https://spec.org/power_ssj2008/results/res2022q4/power_ssj2008-20221204-01204.html ) is 1.81x the performance of best published 2P Intel Xeon Platinum 8490H (16,902 overall ssj_ops/W, 2U, https://spec.org/power_ssj2008/results/res2023q2/power_ssj2008-20230507-01251.html ). SPEC® and SPECpower_ssj® are registered trademarks of the Standard Performance Evaluation Corporation. See www.spec.org for more information.

SP5-013D: SPECrate®2017_int_base comparison based on published scores from www.spec.org as of 06/2/2023. Comparison of published 2P AMD EPYC 9654 (1800 SPECrate®2017_int_base, 720 Total TDP W, $23,610 total 1Ku, 192 Total Cores, 2.500 Perf/W, 0.076 Perf/CPU$, http://spec.org/cpu2017/results/res2023q2/cpu2017-20230424-36017.html) is 1.80x the performance of published 2P Intel Xeon Platinum 8490H (1000 SPECrate®2017_int_base, 700 Total TDP W, $34,000 total 1Ku, 120 Total Cores, 1.429 Perf/W, 0.029 Perf/CPU$, http://spec.org/cpu2017/results/res2023q1/cpu2017-20230310-34562.html) [at 1.75x the performance/W] [at 2.59x the performance/CPU$]. Published 2P AMD EPYC 7763 (861 SPECrate®2017_int_base, 560 Total TDP W, $15,780 total 1Ku, 128 Total Cores, 1.538 Perf/W, 0.055 Perf/CPU$, http://spec.org/cpu2017/results/res2021q4/cpu2017-20211121-30148.html) is shown for reference at 0.86x the performance [at 1.08x the performance/W] [at 1.86x the performance/CPU$]. AMD 1Ku pricing and Intel ARK.intel.com specifications and pricing as of 6/1/23. SPEC®, SPEC CPU®, and SPECrate® are registered trademarks of the Standard Performance Evaluation Corporation. See www.spec.org for more information.

SP5-049C: VMmark® 3.1.1 matched pair comparison based on published results as of 6/13/2023. Configurations: 2-node, 2P 96-core EPYC 9654 powered server running VMware ESXi 8.0b (40.66 @ 42 tiles/798 VMs, https://www.vmware.com/content/dam/digitalmarketing/vmware/en/pdf/vmmark/2023-06-13-Lenovo-ThinkSystem-SR665V3.pdf) versus 2-node, 2P 60-core Xeon Platinum 8490H running VMware ESXi 8.0 GA (23.38 @ 23 tiles/437 VMs, https://www.vmware.com/content/dam/digitalmarketing/vmware/en/pdf/vmmark/2023-03-21-Fujitsu-PRIMERGY-RX2540M7.pdf) for 1.74x the score and 1.75x the tile (VM) capacity. 2-node, 2P EPYC 7763-powered server (23.33 @ 24 tiles/456 VMs, https://www.vmware.com/content/dam/digitalmarketing/vmware/en/pdf/vmmark/2022-02-08-Fujitsu-RX2450M1.pdf) shown at 0.98x performance for reference. VMmark is a registered trademark of VMware in the US or other countries.

SP5-051: TPCx-AI SF3 derivative workload comparison based on AMD internal testing running multiple VM instances as of 6/13/2023. The aggregate end-to-end AI throughput test is derived from the TPCx-AI benchmark and as such is not comparable to published TPCx-AI results, as the end-to-end AI throughput test results do not comply with the TPCx-AI Specification. Configurations: 2 x AMD EPYC 9754 on Titanite (BIOS and Settings: AMI Core Ver. 5.25, Project Ver. RTI1000F and Default BIOS settings (SMT=on,  Determinism=Auto, NPS=1)), 1.5TB (24) Dual-Rank DDR5-4800 64GB DIMMs, 1DPC, SK Hynix SHGP31-500GM 500GB NVMe, Ubuntu® 22.04 LTS (8-instances, 30 vCPUs/instance, 1841 AI test cases/min); 2 x AMD EPYC 9654 on Titanite (BIOS and Settings: AMI Core Ver. 5.25, Project Ver. RTI1000F and Default BIOS settings (SMT=on, Determinism=Auto, NPS=1)), 1.5TB (24) Dual-Rank DDR5-4800 64GB DIMMs, 1DPC, Samsung SSD 983 DCT 960GB, Ubuntu 22.04.1 LTS (6-instance, 28 vCPUs/instance, 1554 AI test cases/min); 2 x Intel(R) Xeon(R) Platinum 8490H on Dell PowerEdge R760 (BIOS and Settings: ESE110Q-1.10 and Package C1E, Default BIOS settings (C State=Disabled, Hyper-Threading, Turbo boost= enabled (ALL)=Enabled, SNC (Sub NUMA)=Disabled)), 2TB (32) Dual-Rank DDR5-4800 64GB DIMMs, 1DPC, Dell 1.7TB NVMe, Ubuntu 22.04.2 LTS (4-instance, 30 vCPUs/instance, 831 AI test cases/min). Results may vary due to factors including system configurations, software versions and BIOS settings. TPC Benchmark is a trademark of the TPC.

SP5-056B: SAP® SD 2-tier comparison based on published results as of 6/13/2023. Configurations: 2P 96-core EPYC 9654 powered server (148,000 benchmark users, https://www.sap.com/dmc/benchmark/2022/Cert22029.pdf) versus 2P 60-core Xeon Platinum 8490H (77,105 benchmark users, https://www.sap.com/dmc/benchmark/2023/Cert23021.pdf) for 1.92x the number of SAP SD benchmark users. 2P EPYC 7763 powered server (75,000 benchmark users, https://www.sap.com/dmc/benchmark/2021/Cert21021.pdf) shown at 0.98x the performance for reference. For more details see http://www.sap.com/benchmark. SAP and SAP logo are the trademarks or registered trademarks of SAP SE (or an SAP affiliate company) in Germany and in several other countries.

SP5-104A: SPECjbb® 2015-MultiJVM Critical based on published scores from www.spec.org as of 3/31/2023. Configurations: 2P AMD EPYC 9654 (664,375 SPECjbb®2015 MultiJVM max-jOPS, 622,315 SPECjbb®2015 MultiJVM critical-jOPS, 192 Total Cores, https://www.spec.org/jbb2015/results/res2022q4/jbb2015-20221019-00860.html) is 1.69x the critical-jOPS performance of published 2P Intel Xeon Platinum 8490H (458,295 SPECjbb®2015 MultiJVM max-jOPS, 368,979 SPECjbb®2015 MultiJVM critical-jOPS, 120 Total Cores, http://www.spec.org/jbb2015/results/res2023q1/jbb2015-20230119-01007.html).

SP5-149: Container density throughput based on sustaining ~25k e-commerce Java Ops/sec/container until exceeding SLA utilizing >90% of the total cores on composite server-side Java workload as measured by AMD as of 6/13/2023. Common container settings: allocated 40GB memory, similar disks & NICs.  2P server configurations: 2P EPYC 9754 128C/256T SMT ON, Memory: 1.5TB = 24 x 64 GB DDR5 4800, OS Ubuntu 22.04, NPS Setting: L3 as NUMA running 16 vCPUs vs. 2P Xeon Platinum 8490H 60C/120T HT ON, Memory: 2TB = 32 x 64 GB DDR5 4800, OS Ubuntu 22.04, NPS Setting: NPS 2 running 16 vCPUs vs. 2P Ampere Altra Max 128-30, Memory: 1TB =16 x 64GB DDR3200, OS Ubuntu 22.04, NPS Setting: NPS 1 running 25C. Results may vary due to factors including system configurations, software versions and BIOS settings.

MI300-005: Calculations conducted by AMD Performance Labs as of May 17, 2023, for the AMD Instinct™ MI300X OAM accelerator 750W (192 GB HBM3) designed with AMD CDNA™ 3 5nm FinFet process technology resulted in 192 GB HBM3 memory capacity and 5.218 TFLOPS sustained peak memory bandwidth performance. MI300X memory bus interface is 8,192 and memory data rate is 5.6 Gbps for total sustained peak memory bandwidth     of 5.218 TB/s (8,192 bits memory bus interface * 5.6 Gbps memory data rate/8)*0.91 delivered adjustment.  The highest published results on the NVidia Hopper H100 (80GB) SXM GPU accelerator resulted in 80GB HBM3 memory capacity and 3.35 TB/s GPU memory bandwidth performance.

MI300-07K:Measurements by internal AMD Performance Labs as of June 2, 2023 on current specifications and/or internal engineering calculations. Large Language Model (LLM) run or calculated with FP16 precision to determine the minimum number of GPUs needed to run the Falcon (40B parameter) models. Tested result configurations: AMD Lab system consisting of 1x EPYC 9654 (96-core) CPU with 1x AMD Instinct™ MI300X (192GB HBM3, OAM Module) 750W accelerator tested at FP16 precision. Server manufacturers may vary configuration offerings yielding different results.

MI300-08K - Measurements by internal AMD Performance Labs as of June 2, 2023 on current specifications and/or internal engineering calculations. Large Language Model (LLM) run comparisons with FP16 precision to determine the minimum number of GPUs needed to run the Falcon (40B parameters); GPT-3 (175 Billion parameters), PaLM 2 (340 Billion parameters); PaLM (540 Billion parameters) models.Calculated estimates based on GPU-only memory size versus memory required by the model at defined parameters plus 10% overhead.

Calculations rely on published and sometimes preliminary model memory sizes. Tested result configurations: AMD Lab system consisting of 1x EPYC 9654 (96-core) CPU with 1x AMD Instinct™ MI300X (192GB HBM3, OAM Module) 750W accelerator Vs. Competitve testing done on Cirrascale Cloud Services comparable instance with permission. 

Calculated estimates may vary based on final model size; actual and estimates may vary due to actual overhead required and using system memory beyond that of the GPU.  Server manufacturers may vary configuration offerings yielding different results.