Some Power11 performance and efficiency information

I had been unable to find any information comparing the performance and efficiency of the new IBM Power11 to previous IBM Power servers. I reached out for this information to people I know and they shared with me an IBM document, whose information I have copied into this post.

Before we get started I need to briefly explain what rPerf is. It is a method to approximate the difference in performance between two Power servers. rPerf is only for AIX. For IBM i performance CPW is used. I found an IBM page explaining what rPerf is here.

All of the quotes I giving below are from the document I received. They are divided into Performance and IT Efficiency. While the document did not group them together I am doing so. I am not including the disclaimers with the quotes, as that will make it difficult to read. The disclaimers can be found at the bottom of this post.

 

Performance

1. Up to 45% more capacity with higher core counts in entry and mid-range systems compared to Power10

2. Power11 offers up to 55% better core performance compared to Power9

3. Up to 25% improvement in per core performance by leveraging Resource Groups for shared processor configurations on Power E1180 compared to E1080

4. Power E1150 delivers 55% improvement in per core database performance with a 25% better transactional response time compared to Power E950

5. Power E1150 delivers 25% better transactional response time compared to Power E950 at similar throughput

6. 50% more throughput (tps) for IBM FTM High Value Payments with S1122 compared to S1022

7. 2X performance / core for IBM FTM High Value Payments for S1122 versus compared x86

8. 51% lower 3-year solution cost for IBM FTM High Value Payments for S1122 versus compared x86

9. 4X performance / core running Temenos Transact on E1180 versus compared x86

10. 50% more throughput for open source core banking with S1122 compared to S1022 using Apache Fineract

11. 3.5x better performance per core with Power11 versus compared x86 systems

12. Up to 15% improved out-of-box performance in users for a commercial ERP workload with Power E1180 compared to Power E1080

13. Up to 58% latency reduction (minutes/documents) for AI inferencing with S1122 compared to S1022 with the same number of users

14. Up to 7.2X more documents/hour for AI inferencing with S1122 compared to S1022 for the same workload

15. Up to 6.9X better performance/$ for AI inferencing with S1122 compared to S1022 for the same workload

 

IT efficiency

16. 10% more rPerf per Watt with Power E1180 compared to Power E1080

17. 20% more rPerf per Watt with Power E1150 compared to Power E1050

18. 22% more rPerf per Watt with Power S1124 compared to Power S1024

19. 37% more rPerf per Watt with Power S1122 compared to Power S1022

20. Up to 37% more rPerf per Watt with Power11 compared to Power10

21. Up to 67% less energy usage and carbon emissions for Power11 compared to Power9

22. 59% less energy and carbon emissions with Power E1150 compared to Power E950

23. 67% less energy and carbon emissions with Power S1124 compared to Power S924

24. 60% less energy and carbon emissions with Power S1122 compared to Power S922

 

Disclaimers

1. Based upon current IBM Power rPerf and CPW estimates for E1150, S1124 and S1122 versus E1050, S1024 and S1022 respectively.

2. Based upon IBM internal measurements of a commercial core banking solution running on IBM Power E950 compared to an E1150.

3. Based on IBM internal measurements of a commercial transactional database(OLTP) workload consisting of 10 partitions executing on an E1080 (16x12 Power10 CPUs @ 3.6-4.15GHz) compared to the same workload executing on an E1180 (16x12 Power11 CPUs @ 3.9-4.4GHz) within 2 Resource Groups with 4 and 6 partitions.

4. Based on IBM internal measurements for similar throughput of a commercial banking transactional database(OLTP) workload executing on 1xE1150 with 9 Power11 CPUs @ 3.5-4.2GHz compared 1xE950 with 14 Power9 CPUs @ 3.2-3.8GHz

5. Based on IBM internal measurements for similar throughput of a commercial banking transactional database(OLTP) workload executing on 1xE1150 with 9 Power11 CPUs @ 3.5-4.2GHz compared 1xE950 with 14 Power9 CPUs @ 3.2-3.8GHz

6. Based on IBM internal testing for FTM v4.0.6 Check data ingest. Transactions per second were measured for Red Hat OpenShift Container Platform 4.16 worker nodes running IBM FTM v4.0.6 High Value Payments. Results valid as of May 29, 2025 and conducted under laboratory conditions, individual results can vary based on workload size, use of storage subsystems and other conditions. Comparison is based on an IBM Power S1122 (2x30-core/1TB) with 45 worker cores and 10 Db2 cores versus S1022 (2x20-core/1TB) with 30 worker cores and 4 Db2 cores. Tests were run with IBM FTM High Value Payments accessing Db2 databases on AIX.

7. Based on IBM internal testing for FTM v4.0.6 Check data ingest. Transactions per second were measured for Red Hat OpenShift Container Platform 4.16 worker nodes running FTM v4.0.6 High Value Payments. Results valid as of May 29, 2025 and conducted under laboratory conditions, individual results can vary based on workload size, use of storage subsystems and other conditions. Comparison is based on an IBM Power S1122 (2x30-core/1TB) with 45 worker cores and 10 Db2 cores versus Intel Xeon x86 (2x72-core/1TB) with 108 worker cores and 22 Db2 cores. Tests were run with FTM High Value Payments accessing Db2 databases on AIX on a Power S1122 system and on RHEL on Intel Xeon server.

8. Based on IBM internal testing for FTM v4.0.6 Check data ingest. Transactions per second were measured for Red Hat OpenShift Container Platform 4.16 worker nodes running FTM v4.0.6 High Value Payments. Results valid as of May 29, 2025 and conducted under laboratory conditions, individual results can vary based on workload size, use of storage subsystems and other conditions. Comparison is based on an IBM Power S1122 (2x30-core/1TB) with 45 worker cores and 10 Db2 cores versus Intel Xeon x86 (2x72-core/1TB) with 108 worker cores and 22 Db2 cores. Tests were run with FTM High Value Payments accessing Db2 databases on AIX on a Power S1122 system and on RHEL on Intel Xeon server. Total cost of ownership (TCO)/solution cost is defined as hardware, software, and maintenance costs over a period of three years and price-performance is performance/3-yr TCO where hardware list pricing is based on IBM Power S1122 https://www.ibm.com/products/ and extrapolated industry standard x86 list pricing based on IBM internal industry knowledge and software list pricing available at https://www.redhat.com/en/technologies/cloud-computing/openshift/pricing.

9. Based on IBM internal testing for Temenos Transact core banking workload (transactions per second) accessing Linux EDB databases each running 2,560,256 transactions using JMeter V5.6.3, the error ratio was kept under 1%, Results valid as of June 25, 2025 and conducted under laboratory conditions, individual results can vary based on workload size, use of storage subsystems and other conditions. Comparison is based on an IBM Power E1180 (16x16 core) with 32 worker cores versus Intel Xeon x86 (4x60 core) with 128 worker cores. Tests were run with Red Hat Enterprise Linux 9.6 on the OpenShift Container Platform Helper, Red Hat OpenShift 4.18.13 and EDB 17.5.0 on both servers.

10. Based on IBM internal testing for open source Apache Fineract® benchmarking. Results valid as of May 31, 2025 and conducted under laboratory conditions, individual results can vary based on workload size, use of storage subsystems and other conditions. Comparison is based on an IBM Power S1122 (2x30-core/2TB) versus S1022 (2x20-core/2TB). Tests were run with Apache Fineract® with Enterprise Database on Red Hat 9.4.

11. Performance is based on Quantitative Performance Index (QPI) data as of May 15, 2025 from IDC available at https://www.idc.com/about/qpi. IBM Power E1150 (4x30c Power11 at 3.0-4.1GHz) QPI of 241,000E versus HPE Compute Scale-up Server 3200 (4x60-core Intel cores at 1.9GHz) QPI of 208,898 and utilizations of 75% for E1150 based on IBM Power Performance Utilization Guarantee and 40% for x86.

12. Based on IBM internal testing of multi-instance SAP NetWeaver running on an E1080 (8x12 core) compared to E1180 (8x12 core) at various utilization levels. Both systems utilized AIX and SAP best practices and no additional performance tuning.

13. Based on IBM internal testing of batched processing and measuring time (minutes) per documents for Red Hat OpenShift Container Platform 4.16 worker nodes running Summarization with Granite3.3-2b-instruct input ~560 tokens, output 200 tokens, batch size=1. Results valid as of June 23, 2025 and conducted under laboratory conditions, individual results can vary based on workload size, use of storage subsystems and other conditions. Comparison is based on an IBM Power S1122 (2x30-core) with 48 worker cores and S1022 (2x20-core) with 32 worker cores.

14. Based on IBM internal testing of batched processing and measuring documents per hour for Red Hat OpenShift Container Platform 4.16 worker nodes running Summarization with Granite3.3-2b-instruct input ~560 tokens, output 200 tokens, batch size=8. Results valid as of June 23, 2025 and conducted under laboratory conditions, individual results can vary based on workload size, use of storage subsystems and other conditions. Comparison is based on an IBM Power S1122 (2x30-core) with 48 worker cores and S1022 (2x20-core) with 32 worker cores.

15. Based on IBM internal testing of batched processing and measuring documents per hour for Red Hat OpenShift Container Platform 4.16 worker nodes running Summarization with Granite3.3-2b-instruct input ~560 tokens, output 200 tokens, batch size=8. Results valid as of June 23, 2025 and conducted under laboratory conditions, individual results can vary based on workload size, use of storage subsystems and other conditions. Comparison is based on an IBM Power S1122 (2x30-core) with 48 worker cores and S1022 (2x20-core) with 32 worker cores.
Pricing is based on solution cost defined as hardware and software costs price-performance is performance/solution cost where hardware list pricing is based on IBM Power S1122 and S1022 price: https://www.ibm.com/products/ and software list pricing is based on Red Hat OpenShift price: https://www.redhat.com/en/technologies/cloud-computing/openshift/pricing.

16. Based on maximum configuration at 100% utilization under typical operating conditions where Power E1080 is 7999 rPerf @ 17,392 Watts (0.460 rPerf/Watt), Power E1180 is 8875E rPerf @ 18,048E Watts (0.492 rPerf/Watt); 0.492 / 0.460 = .1.07 more rPerf/Watt

17. Based on maximum configuration at 100% utilization under typical operating conditions where Power E1050 is 2689 rPerf @ 3,840 Watts (0.700 rPerf/Watt), Power E1150 is 3402E rPerf @ 4,168E Watts (0.816 rPerf/Watt); 0.816 / 0.700 = .1.17 more rPerf/Watt

18. Based on maximum configuration at 100% utilization under typical operating conditions where Power S1024 is 1331 rPerf @ 2,707 Watts (0.492 rPerf/Watt), Power S1124 is 1686E rPerf @ 2,907E Watts (0.580 rPerf/Watt); 0.580 / 0.492 = .1.18 more rPerf/Watt

19. Based on maximum configuration at 100% utilization under typical operating conditions where Power S1022 is 1024 rPerf @ 1,684 Watts (0.608 rPerf/Watt), Power S1124 is 1487E rPerf @ 1,834E Watts (0.811 rPerf/Watt); 0.811 / 0.608 = .1.33 more rPerf/Watt

20. Based on maximum configuration at 100% utilization under typical operating conditions across the all Power11 systems.: Power E1080 is 7999 rPerf @ 17,392 Watts (0.460 rPerf/Watt), Power E1180 is 9141 rPerf @ 18,048E Watts (0.506 rPerf/Watt); 0.506 / 0.460 = .1.10 more rPerf/Watt , Power E1050 is 2689 rPerf @ 3,840 Watts (0.700 rPerf/Watt), Power E1150 is 3504 rPerf @ 4,168E Watts (0.841 rPerf/Watt); 0.842 / 0.700 = .1.20 more rPerf/Watt , Power S1024 is 1331 rPerf @ 2,707 Watts (0.492 rPerf/Watt), Power S1124 is 1737 rPerf @ 2,907E Watts (0.598 rPerf/Watt); 0.598 / 0.492 = .1.22 more rPerf/Watt , Power S1022 is 1024 rPerf @ 1,684 Watts (0.608 rPerf/Watt), Power S1124 is 1532 rPerf @ 1,834E Watts (0.835 rPerf/Watt); 0.835 / 0.608 = .1.37 more rPerf/Watt

21. Based on maximum rPerf ratings consolidating 3 Power9 systems onto a single Power11 system at 100% utilization under typical operating condition: 3 Power E950 (40 cores/1034 rPerf each) @ 3,014 Watts and 1 Power E1150 (120 cores / 3504 rPerf) @ 5,200E Watts; 3 Power S924 (20 cores/500 rPerf each) @ 2,759 Watts and 1 Power S1124 (60 cores / 1737 rPerf) @ 2,750E Watts; 3 Power S922 (20 cores/ 426 rPerf each) @ 1,880 Watts and 1 Power S1122 (60 cores / 1532 rPerf) @ 2240E Watts.

22. Based on maximum rPerf ratings consolidating 3 Power E950 (40 cores/1034 rPerf each) @ 3,014 Watts onto 1 Power E1150 (120 cores / 3504 rPerf) @ 5,200E Watts

23. Based on maximum rPerf ratings consolidating 3 Power S924 (20 cores/500 rPerf each) @ 2,759 Watts and 1 Power S1124 (60 cores / 1737 rPerf) @ 2,750E Watts

24. Based on Based on maximum rPerf ratings consolidating 3 Power S922 (20 cores/426 rPerf each) @ 1,880 Watts and 1 Power S1122 (60 cores / 1532 rPerf) @ 2240E Watts