Up to 2.4x higher summarization performance and 1.68x perf/watt with Intel® Xeon® 6972P [MRDIMM] vs. prior generation (Llama3-8B - int8)

Up to 3.08x higher summarization performance and 2.16x perf/watt with Intel® Xeon® 6972P [MRDIMM] vs. prior generation (Llama2-7B - int8)

Up to 1.86x (int8) and 1.82x (bf16) higher ResNet-50 Real-Time Inference performance with the new Intel® Xeon® 6980P [MRDIMM] vs. prior generation

Up to 1.88x (int8) and 1.93x (bf16) higher BERT-Large batch Inference performance with the new Intel® Xeon® 6980P [MRDIMM] vs. prior generation

Up to 1.83x (int8) and 1.7x (bf16) higher ResNet-50 batch Inference performance with the new Intel® Xeon® 6980P [MRDIMM] vs. prior generation

Up to 16.7x (int8) higher BERT-Large real-time Inference performance with the new Intel® Xeon® 6980P [MRDIMM] vs. Intel Xeon 8280

Up to 13.7x (int8) higher ResNet-50 real-time Inference performance with the new Intel® Xeon® 6980P [MRDIMM] vs. Intel Xeon 8280

Up to 13.6x (int8) higher BERT-Large batch Inference performance with the new Intel® Xeon® 6980P [MRDIMM] vs. Intel Xeon 8280

Up to 17.4x (int8) higher ResNet-50 batch Inference performance with the new Intel® Xeon® 6980P [MRDIMM] vs. Intel Xeon 8280

Up to 14x (int8) higher DLRM batch Inference performance with the new Intel® Xeon® 6980P [MRDIMM] vs. Intel Xeon 8280

Up to 4.29x (int8) higher BERT-Large batch inference performance with the new Intel Xeon 6980P [MRDIMM] vs. AMD EPYC 9654

Up to 5.5x (int8) higher ResNet-50 batch Inference performance with the new Intel® Xeon® 6980P [MRDIMM] vs. AMD EPYC 9654

Up to 3.99x (int8) higher DLRM batch Inference performance with the new Intel® Xeon® 6980P [MRDIMM] vs. AMD EPYC 9654

Up to 1.9x (int8) and 1.83x (bf16) higher Stable Diffusion Inference performance with the new Intel® Xeon® 6980P [MRDIMM] vs. prior generation

Up to 1.72x (int8) and 1.8x (bf16) higher real-time Vision-Transformer performance and 1.6x (int8) and 1.9x (bf16) higher batched Vision-transformer Inference performance with the new Intel® Xeon® 6980P [MRDIMM] vs. prior generation

Up to 2x (int8) and 1.5x (bf16) higher real-time YOLO-v7 performance and 1.61x (int8) and 1.95x (bf16) higher batched YOLO-v7 Inference performance with the new Intel® Xeon® 6980P [MRDIMM] vs. prior generation

Up to 2.3x higher Llama-2-7B training (bf16) performance with the new Intel® Xeon® 6980P vs. prior generation

Up to 3.0x higher batched Llama2-7B inference (int8) with the new Intel® Xeon® 6972P vs. AMD EPYC 9654

Up to 4.0x higher batched LLama3-8B inference (int8) with the new Intel® Xeon® 6972P vs. AMD EPYC 9654

Up to 2.36x higher batched GPT-J-6B LLM Sumarization (1024/128, int8) with the new Intel® Xeon® 6972P vs. AMD EPYC 9654

Up to 2.40x higher batched Llama2-7B LLM Sumarization (1024/128, int8) with the new Intel® Xeon® 6972P vs. AMD EPYC 9654

Up to 3.56x higher batched Llama3-8B LLM Sumarization (1024/128, int8) with the new Intel® Xeon® 6972P vs. AMD EPYC 9654

Up to 2x higher Llama-13B (BF16) performance with Intel® Xeon® 6980P vs. Intel Xeon 8592+.

Up to 7x higher HammerDB MySQL OLTP performance with the new Intel® Xeon® 6980P vs. Intel® Xeon® 8280

Up to 1.72x higher HammerDB MySQL OLTP performance with the new Intel® Xeon® 6980P vs. AMD EPYC 9654.

Up to 1.40x higher HammerDB MySQL OLTP performance with the new Intel® Xeon® 6980P vs. AMD EPYC 9754.

Up to 7x higher HammerDB MySQL OLTP performance with the new Intel® Xeon® 6980P vs. Intel® Xeon® 8280

Up to 1.72x higher HammerDB MySQL OLTP performance with the new Intel® Xeon® 6980P vs. AMD EPYC 9654.

Up to 1.40x higher HammerDB MySQL OLTP performance with the new Intel® Xeon® 6980P vs. AMD EPYC 9754.

Up to 8x higher MongoDB performance with the new Intel® Xeon® 6980P vs. Intel® Xeon® 8280

Up to 2.38x higher MongoDB performance with the new Intel® Xeon® 6980P vs. AMD EPYC 9654

Up to 2.0x higher MongoDB performance with the new Intel® Xeon® 6980P vs. AMD EPYC 9754

Up to 1.72x higher Redis VectorDB Similarity search performance with the new Intel® Xeon® 6980P [MRDIMM] vs. AMD EPYC 9654

Up to 1.67x higher Redis VectorDB Similarity search performance with the new Intel® Xeon® 6980P [MRDIMM] vs. AMD EPYC 9754

Up to 5.6x higher Redis Memtier performance with the new Intel® Xeon® 6980P [DDR] vs. Intel® Xeon® 8280

Up to 1.71x higher Redis Memtier performance with the new Intel® Xeon® 6980P [DDR] vs. AMD EPYC 9654

Up to 1.92x higher Redis Memtier performance with the new Intel® Xeon® 6980P [DDR] vs. AMD EPYC 9754

Up to 1.45x higher CassandraDB performance with the new Intel® Xeon® 6980P [DDR] at 10ms P99 SLA vs. AMD EPYC 9754

Up to 2.06x higher integer throughput performance and up to 1.44x higher integer throughput performance/watt (CPU TDP) with new Intel® Xeon® 6980P [MRDIMM] over prior generation Intel® Xeon® Platinum 8592+

Up to 2.07x higher floating point throughput performance and up to 1.45x higher floating point throughput performance/watt (CPU TDP) with new Intel® Xeon® 6980P [MRDIMM] over prior generation Intel® Xeon® Platinum 8592+

Up to 2.14x higher memory bandwidth and up to 1.5x higher memory bandwidth performance/watt (CPU TDP) with new Intel® Xeon® 6980P [MRDIMM] over prior generation Intel® Xeon® Platinum 8592+

Up to 2.05x higher LINPACK performance and up to 1.44x higher LINPACK performance/watt (CPU TDP) with new Intel® Xeon® 6980P [DDR] over prior generation Intel® Xeon® Platinum 8592+

Up to 1.6x higher performance average per watt vs. 5th Gen Intel Xeon

Up to 6.4x higher floating point throughput with Intel® Xeon® 6980P vs. Intel® Xeon® 8280.

Up to 2x higher integer throughput with Intel® Xeon® 6980P vs. Intel® Xeon® 8592+

Up to 2x higher floating point throughput with Intel® Xeon® 6980P vs. Intel® Xeon® 8592+

Up to 1.63x higher floating point throughput performance on Intel® Xeon® 6980P [MRDIMM] vs. AMD EPYC 9654

Up to 1.74x higher memory bandwidth on Intel® Xeon® 6980P [MRDIMM] vs. AMD EPYC 9654

Up to 2.04x higher LINPACK performance on Intel® Xeon® 6972P [DDR] vs. AMD EPYC 9654

Up to 1.20x higher integer throughput performance on Intel® Xeon® 6980P [MRDIMM] vs. AMD EPYC 9754

Up to 1.59x higher floating point throughput performance on Intel® Xeon® 6980P [MRDIMM] vs. AMD EPYC 9754

Up to 1.61x higher memory bandwidth on Intel® Xeon® 6980P [MRDIMM] vs. AMD EPYC 9754

Up to 1.79x higher LINPACK performance on Intel® Xeon® 6972P [DDR] vs. AMD EPYC 9754

Up to 2.31x higher NEMO performance (geomean 2x) with Intel® Xeon® 6980P [MRDIMM] processor compared to Intel Xeon 8592+

Up to 1.84x MPAS-A (geomean of 2) performance with Intel Xeon 6980P [MRDIMM] vs. AMD EPYC 9654

Up to 1.75x NEMO (geomean of 2) performance with Intel Xeon 6980P [MRDIMM] vs. AMD EPYC 9654

Up to 1.28x WRF (geomean of 2) performance with Intel Xeon 6980P [MRDIMM] vs. AMD EPYC 9654

Up to 1.7x LAMMPS [MRDIMM] performance with Intel® Xeon® 6980P vs. AMD EPYC 9654

Up to 2.26x NAMD [MRDIMM] performance with Intel® Xeon® 6980P vs. AMD EPYC 9654

Up to 2.01x RELION [DDR] performance with Intel® Xeon® 6980P vs. AMD EPYC 9654

Up to 1.54x ANSYS FLUENT [MRDIMM] performance with Intel® Xeon® 6980P vs. AMD EPYC 9654

Up to 1.45x ANSYS LS-DYNA [MRDIMM] performance with Intel® Xeon® 6980P vs. AMD EPYC 9654

Up to 2.05x OpenFOAM [MRDIMM] performance with Intel® Xeon® 6980P vs. AMD EPYC 9654

Up to 2.3x higher HPCG performance with Intel® Xeon® 6980P vs. Intel® Xeon® 8592+.

Up to 2.1x higher OpenFOAM performance with Intel® Xeon® 6980P [MRDIMM] vs. AMD EPYC 9654

Up to 1.49x higher IOPS/watt on Intel® Xeon® 6980P with Intel VROC (8 disk RAID0) vs. legacy architecture (HBA)

This is based on a geomean average of TCO savings for scenarios running NGINX TLS, MongoDB, OpenFOAM, and ResNet50 workloads. See [9T3] ,[9T4], [9T5], [9T6].

Up to 1.8x higher Java Throughput within a given SLA with Intel Xeon 6980P processor vs. Intel® Xeon® 8592+

Up to 6.1x higher Java Throughput with Intel Xeon 6980P processor vs. Intel® Xeon® 8280

Up to 5.7x higher Java Throughput within a given SLA with Intel Xeon 6980P processor vs. Intel® Xeon® 8280

Up to 1.32x higher Java Throughput with the new Intel Xeon 6980P [MRDIMM] vs. AMD EPYC 9654

Up to 1.24x higher Java Throughput with the new Intel Xeon 6980P [MRDIMM] vs. AMD EPYC 9754

Up to 1.42x higher Java Throughput within a given SLA with the new Intel Xeon 6980P [MRDIMM] vs. AMD EPYC 9654

Up to 1.27x higher Java Throughput within a given SLA with the new Intel Xeon 6980P [MRDIMM] vs. AMD EPYC 9754

Up to 1.5x higher Java Throughput with Intel Xeon 6972P processor vs. Intel® Xeon® 8592+

Up to 1.56x higher Java Throughput within a given SLA with Intel Xeon 6972P processor vs. Intel® Xeon® 8592+

Up to 5.37x higher Java Throughput with Intel Xeon 6972P processor vs. Intel® Xeon® 8280

Up to 4.86x higher Java Throughput within a given SLA with Intel Xeon 6972P processor vs. Intel® Xeon® 8280

Up to 7x higher NGINX performance with the new Intel® Xeon® 6980P vs. Intel® Xeon® 8280

Up to 1.84x higher NGINX performance with the new Intel® Xeon® 6972P vs. Intel® Xeon® 8592+

Up to 6.4x higher NGINX performance with the new Intel® Xeon® 6972P vs. Intel® Xeon® 8280

Up to 1.89x higher throughput performance on SVT-HEVC Media Transcode with Intel® Xeon® 6980P vs. Intel Xeon Platinum 8592+

Up to 2.14x higher throughput performance on AVC Media Transcode with Intel® Xeon® 6980P vs. Intel Xeon Platinum 8592+

Up to 2.19x higher throughput performance on x265 Media Transcode with Intel® Xeon® 6980P vs. Intel Xeon Platinum 8592+

Up to 1.30x higher throughput performance on SVT-HEVC Media Transcode with the new Intel Xeon 6980P [MRDIMM] vs. AMD EPYC 9654

Up to 1.08x higher throughput performance on SVT-HEVC Media Transcode with the new Intel Xeon 6980P [MRDIMM] vs. AMD EPYC 9754

Up to 6x higher WordPress performance with Intel Xeon 6980P [DDR] processor vs. prior generation 8280

Up to 1.37x higher WordPress performance with the new Intel Xeon 6980P [DDR] vs. AMD EPYC 9654

Up to 1.15x higher WordPress performance with the new Intel Xeon 6980P [DDR] vs. AMD EPYC 9754

Up to 3.21x higher NGINX TLS throughput performance with the new Intel Xeon 6980P [MRDIMM] vs. AMD EPYC 9754

Intel® Xeon® 6 with P-cores delivers up to 58% lower total cost of ownership (TCO) than AMD EPYC 9654 based servers running a NGINX TLS (1S) workload.

Dec 2024 update:

Use 10 racks (140 servers) of Intel® Xeon® 6972P based servers running NGINX (1S) instead of 21 racks (420 servers) of AMD EPYC 9654 based servers and save 8130 MWh of energy, reduce carbon footprint by 3447 metric tons CO2, and save $11799K in total cost of ownership over 4-years. Intel Xeon 6972P delivers 3.04x higher performance and 2.06x higher performance/watt per server vs Intel Xeon AMD EPYC 9654 on NGINX TLS (1S).

For 21 racks of AMD EPYC 9654 based servers over 4-years, estimated as of Sep 2024: CapEx costs: $13911k, OpEx costs (4 year, includes power and cooling utility costs, infrastructure and hardware maintenance costs): $6420k, Energy use: 15988MWh, CO2 emissions: 6778 metric tons.

For 10 rack of Intel Xeon 6972P based servers over 4-years, estimated as of Dec 2024: CapEx costs: $6063k, OpEx costs (4 year, includes power and cooling utility costs, infrastructure and hardware maintenance costs): $2468k: Energy use: 7857MWh, CO2 emissions: 3331 metric tons.

Costs based on Intel estimates and information from thinkmate.com as of Dec 2024.

Sep 2024:

Intel® Xeon® 6 with P-cores delivers up to 49% lower total cost of ownership (TCO) than AMD EPYC 9654 based servers running a NGINX TLS (1S) workload. Intel® Xeon® 6 with P-cores delivers up to 49% lower total cost of ownership (TCO) than AMD EPYC 9654 based servers running a NGINX TLS (1S) workload. Use 10 racks (140 servers) of Intel® Xeon® 6972P based servers running NGINX (1S) instead of 21 racks (420 servers) of AMD EPYC 9654 based servers and save 8130 MWh of energy, reduce carbon footprint by 3447 metric tons CO2, and save $9960K in total cost of ownership over 4-years. Intel Xeon 6972P delivers 3.04x higher performance and 2.06x higher performance/watt per server vs Intel Xeon AMD EPYC 9654 on NGINX TLS (1S). 6972P: 1-node, pre-production platform, 2x (1x used) Intel(R) 6972P, 96 cores, HT On, Turbo Off, Integrated Total Memory 1536GB (24x64GB DDR5 6400 MT/s [6400 MT/s]), BIOS BHSDCRB1.IPC.0033.D57.2406240014, microcode 0x11000290, 2x Ethernet Controller 10-Gigabit X540-AT2, 2x Ethernet Controller E810-C for QSFP, 1x 1.7T SAMSUNG MZQL21T9HCJR-00A07, Ubuntu 22.04.4 LTS, 6.5.0-21-generic, Test by Intel as of 07/25/24. AMD EPYC 9654: 1-node, 2x (1x used) AMD EPYC 9654, 96 cores, SMT On, Boost Off, Total Memory 1536GB (24x64GB DDR5 4800 MT/s [4800 MT/s]), BIOS 1.6, microcode 0xa10113e, 2x Ethernet Controller 10G X550T, 2x Ethernet Controller E810-C for QSFP, 1x 3.5T SAMSUNG MZWLJ3T8HBLS-00007, Ubuntu 22.04.4 LTS, 6.5.0-26-generic, Test by Intel as of July 2024. Software: NGINX,NGINX Async v0.5.1, OpenSSL 3.1.3, Intel=QAT driver: QAT _​​Engine v 1.6.1,QAT20.L.1.2.30-00020, AMD=OOB, IPP Crypto 2021.8,IPsec MB v 1.4, Assumptions: 1x 42U, 15KW Rack, supporting 20x 2U rack servers, 1x TOR switch, 1.6 PUE, kWh to kg CO2 factor 0.42394. For 21 racks of AMD EPYC 9654 based servers over 4-years, estimated as of Sep 2024: CapEx costs: $13863k, OpEx costs (4 year, includes power and cooling utility costs, infrastructure and hardware maintenance costs): $6420k, Energy use: 15988MWh, CO2 emissions: 6778 metric tons. For 10 rack of Intel Xeon 6972P based servers over 4-years, estimated as of Sep 2024: CapEx costs: $7854k, OpEx costs (4 year, includes power and cooling utility costs, infrastructure and hardware maintenance costs): $2468k: Energy use: 7857MWh, CO2 emissions: 3331 metric tons. Costs based on Intel estimates and information from thinkmate.com as of Sep 2024. Results may vary.

Intel® Xeon® 6 with P-cores delivers up to 38% lower total cost of ownership (TCO) than AMD EPYC 9654 based servers running a MongoDB (1S) workload.

Dec 2024 update:

Use 10 racks (110 servers) of Intel® Xeon® 6972P based servers running MongoDB (1S) instead of 12 racks (228 servers) of AMD EPYC 9654 based servers and save 1897 MWh of energy, reduce carbon footprint by 804 metric tons CO2, and save $4254K in total cost of ownership over 4-years. Intel Xeon 6972P delivers 2x higher performance and 1.2x higher performance/watt per server vs Intel Xeon AMD EPYC 9654 on MongoDB (1S).

For 12 racks of AMD EPYC 9654 based servers over 4-years, estimated as of Sep 2024: CapEx costs: $7552k, OpEx costs (4 year, includes power and cooling utility costs, infrastructure and hardware maintenance costs): $3611k, Energy use: 9648MWh, CO2 emissions: 4090 metric tons.

For 10 rack of Intel Xeon 6972P based servers over 4-years, estimated as of Dec 2024: CapEx costs: $4764k, OpEx costs (4 year, includes power and cooling utility costs, infrastructure and hardware maintenance costs): $2145k: Energy use: 7750MWh, CO2 emissions: 3286 metric tons.

Costs based on Intel estimates and information from thinkmate.com as of Dec 2024.

Sep 2024:

Intel® Xeon® 6 with P-cores delivers up to 25% lower total cost of ownership (TCO) than AMD EPYC 9654 based servers running a MongoDB (1S) workload. Use 10 racks (110 servers) of Intel® Xeon® 6972P based servers running MongoDB (1S) instead of 12 racks (228 servers) of AMD EPYC 9654 based servers and save 1897 MWh of energy, reduce carbon footprint by 804 metric tons CO2, and save $2821K in total cost of ownership over 4-years. Intel Xeon 6972P delivers 2x higher performance and 1.2x higher performance/watt per server vs Intel Xeon AMD EPYC 9654 on MongoDB (1S). 6972P : 1-node, pre-production platform, 2x (1x used) Intel(R) Xeon(R) 6972P, 96 cores, HT On, Turbo On, Total Memory 1536GB (24x64GB DDR5 6400 MT/s [6400 MT/s]), BIOS BHSDCRB1.IPC.0033.D57.2406240014, microcode 0x11000290, 2x Ethernet Controller 10-Gigabit X540-AT2, 2x Ethernet Controller E810-C for QSFP, 1x 1.7T Micron_​​7400_​​MTFDKCC1T9TDZ, 4x 3.5T KIOXIA KCD8XPUG3T84, Ubuntu 22.04.4 LTS, 6.5.0-21-generic. Test by Intel as of 07/16/24. AMD EPYC 9654: 1-node, 2x (1x used) AMD EPYC 9654 96-Core Processor, 96 cores, SMT On, Boost On, NPS 1, Total Memory 1536GB (24x64GB DDR5 4800 MT/s [4800 MT/s]), BIOS 1.6, microcode 0xa10113e, 2x Ethernet Controller E810-C for QSFP, 2x Ethernet Controller 10G X550T, 1x 1.7T SAMSUNG MZQL21T9HCJR-00A07, Ubuntu 22.04.4 LTS, 6.5.0-21-generic. Test by Intel as of 07/29/24. Software: MongoDB 6.0.4, YCSB, multi-instance, read/update 90/10. Assumptions: 1x 42U, 15KW Rack, supporting 20x 2U rack servers, 1x TOR switch, 1.6 PUE, kWh to kg CO2 factor 0.42394. For 12 racks of AMD EPYC 9654 based servers over 4-years, estimated as of Sep 2024: CapEx costs: $7526k, OpEx costs (4 year, includes power and cooling utility costs, infrastructure and hardware maintenance costs): $3611k, Energy use: 9648MWh, CO2 emissions: 4090 metric tons. For 10 rack of Intel Xeon 6972P based servers over 4-years, estimated as of Sep 2024: CapEx costs: $6171k, OpEx costs (4 year, includes power and cooling utility costs, infrastructure and hardware maintenance costs): $2145k: Energy use: 7750MWh, CO2 emissions: 3286 metric tons. Costs based on Intel estimates and information from thinkmate.com as of Sep 2024.

Intel® Xeon® 6 with P-cores delivers up to 29% lower total cost of ownership (TCO) than AMD EPYC 9654 based servers running a OpenFOAM geomean (2) workload.

Dec 2024 update:

Use 10 racks (80 servers) of Intel® Xeon® 6972P based servers running OpenFOAM geomean (2) instead of 11 racks (154 servers) of AMD EPYC 9654 based servers and save 992 MWh of energy, reduce carbon footprint by 421 metric tons CO2, and save $2303K in total cost of ownership over 4-years. Intel Xeon 6972P delivers 1.82x higher performance and 1.06x higher performance/watt per server vs Intel Xeon AMD EPYC 9654 on OpenFOAM geomean (2).

For 11 racks of AMD EPYC 9654 based servers over 4-years, estimated as of Sep 2024: CapEx costs: $4995k, OpEx costs (4 year, includes power and cooling utility costs, infrastructure and hardware maintenance costs): $2770k, Energy use: 9066MWh, CO2 emissions: 3843 metric tons.

For 10 rack of Intel Xeon 6972P based servers over 4-years, estimated as of Dec 2024: CapEx costs: $3586k, OpEx costs (4 year, includes power and cooling utility costs, infrastructure and hardware maintenance costs): $1876k: Energy use: 8073MWh, CO2 emissions: 3423 metric tons.

Costs based on Intel estimates and information from thinkmate.com as of Dec 2024.

Sep 2024:

Intel® Xeon® 6 with P-cores delivers up to 18% lower total cost of ownership (TCO) than AMD EPYC 9654 based servers running a OpenFOAM geomean (2) workload. Use 10 racks (80 servers) of Intel® Xeon® 6972P based servers running OpenFOAM geomean (2) instead of 11 racks (154 servers) of AMD EPYC 9654 based servers and save 992 MWh of energy, reduce carbon footprint by 421 metric tons CO2, and save $1437K in total cost of ownership over 4-years. Intel Xeon 6972P delivers 1.89x higher performance and 1.1x higher performance/watt per server vs Intel Xeon AMD EPYC 9654 on OpenFOAM geomean (2). OpenFOAM (motorbike-20m, motorbike-42m) 6972P w/MRDIMM: Test by Intel as of July 2024, 1 node, 2x Intel Xeon 6972P, HT On, Turbo On, SNC3, 1536 GB MRDIMM-8800, BIOS BHSDREL1.86B.0033.D40.2406180419, ucode=0x11000280, Ubuntu 24.04, Kernel 6.8.0, OpenFOAM v2312 9654: Test by Intel as of April 2024, 1 node, 2x AMD EPYC 9654, SMT On, Boost On, NPS4, 1536 GB DDR5-4800, BIOS 1.9, ucode 0xa101148, Ubuntu 24.04, Kernel 6.8.0, OpenFOAM v2312 Assumptions: 1x 42U, 15KW Rack, supporting 20x 2U rack servers, 1x TOR switch, 1.6 PUE, kWh to kg CO2 factor 0.42394. For 11 racks of AMD EPYC 9654 based servers over 4-years, estimated as of Sep 2024: CapEx costs: $4977k, OpEx costs (4 year, includes power and cooling utility costs, infrastructure and hardware maintenance costs): $2770k, Energy use: 9066MWh, CO2 emissions: 3843 metric tons. For 10 rack of Intel Xeon 6972P based servers over 4-years, estimated as of Sep 2024: CapEx costs: $4433k, OpEx costs (4 year, includes power and cooling utility costs, infrastructure and hardware maintenance costs): $1876k: Energy use: 8073MWh, CO2 emissions: 3423 metric tons. Costs based on Intel estimates and information from thinkmate.com as of Sep 2024. This offering is not approved or endorsed by OpenCFD Limited, producer and distributor of the OpenFOAM software via www.openfoam.com, and owner of the OPENFOAM® and OpenCFD® trademark

Intel® Xeon® 6 with P-cores delivers up to 74% lower total cost of ownership (TCO) than AMD EPYC 9654 based servers running a ResNet50 workload.

Dec 2024 update:

Use 10 racks (100 servers) of Intel® Xeon® 6972P based servers running ResNet50 instead of 31 racks (496 servers) of AMD EPYC 9654 based servers and save 16717 MWh of energy, reduce carbon footprint by 7087 metric tons CO2, and save $18870K in total cost of ownership over 4-years. Intel Xeon 6972P delivers 4.91x higher performance and 2.97x higher performance/watt per server vs Intel Xeon AMD EPYC 9654 on ResNet50.

For 31 racks of AMD EPYC 9654 based servers over 4-years, estimated as of Sep 2024: CapEx costs: $17086k, OpEx costs (4 year, includes power and cooling utility costs, infrastructure and hardware maintenance costs): $8387k, Energy use: 25083MWh, CO2 emissions: 10634 metric tons.

For 10 rack of Intel Xeon 6972P based servers over 4-years, estimated as of Dec 2024: CapEx costs: $4482k, OpEx costs (4 year, includes power and cooling utility costs, infrastructure and hardware maintenance costs): $2121k: Energy use: 8365MWh, CO2 emissions: 3546 metric tons.

Costs based on Intel estimates and information from thinkmate.com as of Dec 2024.

Sep 2024:

Intel® Xeon® 6 with P-cores delivers up to 69% lower total cost of ownership (TCO) than AMD EPYC 9654 based servers running a ResNet50 workload. Use 10 racks (100 servers) of Intel® Xeon® 6972P based servers running ResNet50 instead of 31 racks (496 servers) of AMD EPYC 9654 based servers and save 16717 MWh of energy, reduce carbon footprint by 7087 metric tons CO2, and save $17619K in total cost of ownership over 4-years. Intel Xeon 6972P delivers 4.91x higher performance and 2.97x higher performance/watt per server vs Intel Xeon AMD EPYC 9654 on ResNet50. ResNet50 : Intel Xeon 6972P, MRDIMM: 1-node, pre-production platform, 2x Intel(R) Xeon(R) 6972P, 96 cores, 500W TDP, HT On, Turbo On, Total Memory 1536GB (24x64GB MRDIMM 8800MT/s [8800 MT/s]), BIOS BHSDCRB1.IPC.0033.D57.2406240014, microcode 0x11000290, 2x Dual-Media 10G RDMA Ethernet Controller, 2x Ethernet Controller E810-C for QSFP, 1x 1.7T SAMSUNG MZQL21T9HCJR-00A07, Ubuntu 24.04 LTS, 6.8.0-31-generic. Test by Intel as of 08/05/24. AMD EPYC 9654: 1-node, 2x AMD EPYC 9654, 96 cores, 360W TDP, SMT On, Boost On, NPS 1, Total Memory 1536GB (24x64GB DDR5 4800 MT/s [4800 MT/s]), BIOS 1.6, microcode 0xa101144, 2x Ethernet Controller X550, 1x 3.5T SAMSUNG MZWLJ3T8HBLS-00007, Ubuntu 24.04 LTS, 6.8.0-31-generic. Test by Intel as of 07/09/24. Software: ResNet50 v1.5, Inference: int8, bs=1 (sla=15ms), bs=x, Dataset: ImageNet, Framework: PyTorch:2.4.0, IPEX: 2.4.0, OneDNN: v3.4.2, Modelzoo: https://github.com/intel/ai-reference-models Assumptions: 1x 42U, 15KW Rack, supporting 20x 2U rack servers, 1x TOR switch, 1.6 PUE, kWh to kg CO2 factor 0.42394. For 31 racks of AMD EPYC 9654 based servers over 4-years, estimated as of Sep 2024: CapEx costs: $17026k, OpEx costs (4 year, includes power and cooling utility costs, infrastructure and hardware maintenance costs): $8387k, Energy use: 25083MWh, CO2 emissions: 10634 metric tons. For 10 rack of Intel Xeon 6972P based servers over 4-years, estimated as of Sep 2024: CapEx costs: $5673k, OpEx costs (4 year, includes power and cooling utility costs, infrastructure and hardware maintenance costs): $2121k: Energy use: 8365MWh, CO2 emissions: 3546 metric tons. Costs based on Intel estimates and information from thinkmate.com as of Sep 2024.

Intel® Xeon® 6 with P-cores delivers up to 57% lower total cost of ownership (TCO) than AMD EPYC 9654 based servers running a Llama-2 7B Chatbot (128/128) int8/BSX workload.

Dec 2024 update:

Use 10 racks (80 servers) of Intel® Xeon® 6972P based servers running Llama-2 7B Chatbot (128/128) int8/BSX instead of 18 racks (252 servers) of AMD EPYC 9654 based servers and save 6997 MWh of energy, reduce carbon footprint by 2967 metric tons CO2, and save $7275K in total cost of ownership over 4-years. Intel Xeon 6972P delivers 3.06x higher performance and 1.86x higher performance/watt per server vs Intel Xeon AMD EPYC 9654 on Llama-2 7B Chatbot (128/128) int8/BSX.

For 18 racks of AMD EPYC 9654 based servers over 4-years, estimated as of Sep 2024: CapEx costs: $8174k, OpEx costs (4 year, includes power and cooling utility costs, infrastructure and hardware maintenance costs): $4506k, Energy use: 14622MWh, CO2 emissions: 6199 metric tons.

For 10 rack of Intel Xeon 6972P based servers over 4-years, estimated as of Dec 2024: CapEx costs: $3586k, OpEx costs (4 year, includes power and cooling utility costs, infrastructure and hardware maintenance costs): $1818k: Energy use: 7625MWh, CO2 emissions: 3232 metric tons.

Costs based on Intel estimates and information from thinkmate.com as of Dec 2024.

Sep 2024:

Intel® Xeon® 6 with P-cores delivers up to 49% lower total cost of ownership (TCO) than AMD EPYC 9654 based servers running a Llama-2 7B Chatbot (128/128) int8/BSX workload. Use 10 racks (80 servers) of Intel® Xeon® 6972P based servers running Llama-2 7B Chatbot (128/128) int8/BSX instead of 18 racks (252 servers) of AMD EPYC 9654 based servers and save 6997 MWh of energy, reduce carbon footprint by 2967 metric tons CO2, and save $6292K in total cost of ownership over 4-years. Intel Xeon 6972P delivers 3.06x higher performance and 1.86x higher performance/watt per server vs Intel Xeon AMD EPYC 9654 on Llama-2 7B Chatbot (128/128) int8/BSX. Llama-2 7B Chatbot (128/128) int8/BSX. Intel Xeon 6972P, MRDIMM: 1-node, pre-production platform, 2x Intel(R) Xeon(R) 6972P, 96 cores, 500W TDP, HT On, Turbo On, Total Memory 1536GB (24x64GB MRDIMM 8800MT/s [8800 MT/s]), BIOS BHSDCRB1.IPC.0033.D57.2406240014, microcode 0x11000290, 2x Ethernet Controller 10-Gigabit X540-AT2, 1x 1.7T SAMSUNG MZQL21T9HCJR-00A07, Ubuntu 24.04 LTS, 6.8.0-31-generic. Test by Intel as of 07/10/24. AMD EPYC 9654: 1-node, 2x AMD EPYC 9654, 96 cores, 360W TDP, SMT On, Boost On, NPS 1, Total Memory 1536GB (24x64GB DDR5 4800 MT/s [4800 MT/s]), BIOS 1.6, microcode 0xa101144, 2x Ethernet Controller X550, 1x 3.5T SAMSUNG MZWLJ3T8HBLS-00007, Ubuntu 24.04 LTS, 6.8.0-31-generic. Test by Intel as of 07/09/24. Software: Llama2 7B: int8, Framework: PyTorch:2.4.0, IPEX: 2.4.0, OneDNN: v3.4.2, Modelzoo: https://github.com/intel/ai-reference-models, Model instances=2, bs=16 (128/128), bs=2 (2048/2048), bs=7 (1024/128) Assumptions: 1x 42U, 15KW Rack, supporting 20x 2U rack servers, 1x TOR switch, 1.6 PUE, kWh to kg CO2 factor 0.42394. For 18 racks of AMD EPYC 9654 based servers over 4-years, estimated as of Sep 2024: CapEx costs: $8143k, OpEx costs (4 year, includes power and cooling utility costs, infrastructure and hardware maintenance costs): $4506k, Energy use: 14622MWh, CO2 emissions: 6199 metric tons. For 10 rack of Intel Xeon 6972P based servers over 4-years, estimated as of Sep 2024: CapEx costs: $4539k, OpEx costs (4 year, includes power and cooling utility costs, infrastructure and hardware maintenance costs): $1818k: Energy use: 7625MWh, CO2 emissions: 3232 metric tons. Costs based on Intel estimates and information from thinkmate.com as of Sep 2024.