Commercial Processing Workload (CPW)
IBM's relative capacity rating for processors running IBM i, and the number that shows how much transaction headroom a Power server has left. IBM quotes it for the largest IBM i partition a machine allows, which on several Power 11 processor options is capped at 4 cores.
What does CPW tell you?
CPW is a size label for IBM i work. Think of it like the weight rating on a pickup truck. It does not tell you how fast the truck goes. It tells you roughly how much load it can carry before it struggles. CPW does the same job for the work IBM i shops actually run: lots of short jobs hitting a database at the same time.
IBM names three things the CPW rating is
- A database testA range of database applications, with updates and queries of mixed difficulty, running under commitment control and journaling.
- A crowd testMany users at once, all running the same group of programs against the same data.
- A steady-state approximationA reasonable stand-in for how a normal, database-heavy business application performs compared with another system.
Wording follows the "What the CPW rating is" list in the IBM i Performance Capabilities Reference.
What does CPW not tell you?
IBM is unusually blunt about the limits. CPW is built from one workload IBM keeps in-house, so it is a comparison tool, not a promise about your software.
- It is not an indication of how a system will perform for any one customer situation.
- It is not a test of ad-hoc query performance.
- It is not a test of single-threaded batch throughput, so it will not tell you whether a nightly job finishes inside a 4 hour window.
- It cannot represent the execution speed of any single thread. A capacity number and a speed number are different things.
IBM's own words: the CPW rating is "rigidly defined as a relative capacity metric for rough model comparisons and relative CPU consumption" and is "NOT representative of any specific environment". For sizing a real workload IBM points you at the IBM Systems Workload Estimator instead.
Why does a bigger server sometimes show less CPW?
This is the trap worth knowing before you buy. IBM publishes CPW next to a column called LPAR size, which is the number of cores the IBM i partition was given. On the IBM Power S1122, only the 8-core processor option lets an IBM i partition use all 8 cores. Every larger processor option holds the IBM i partition to 4 cores. So the 8-core machine carries nearly twice the IBM i CPW of the 20-core machine, even though the 20-core machine has more silicon in it.
IBM i CPW on the IBM Power S1122, by processor option
Figures from Section 3a of the IBM Power11 Performance Report, 8 July 2025. The extra cores on the larger options are not wasted; they are available to AIX and Linux partitions on the same machine, including a Linux partition driving accelerator cards. IBM quotes CPW on maximum configurations and supplies all performance estimates as is, with no warranty.
Does CPW measure AI inference?
No, and this is where people get tripped up. CPW rates how much IBM i database and transaction work a processor can carry. An AI model answering a question is a different kind of work on different silicon, and IBM rates that in trillions of operations per second. On IBM Power the Spyre Accelerator sits in a Red Hat Enterprise Linux partition, not in IBM i, so the model's speed never appears in the server's CPW figure at all.
CPW is still worth checking before an AI project, just for a different reason. It tells you whether the IBM i side has capacity left to serve the data a model keeps asking for. It does not tell you how fast the model will answer.
How is CPW different from rPerf?
The same IBM report carries both numbers, which is why they get mixed up. They are separate scales for separate operating systems.
CPW and rPerf side by side
| CPW | rPerf | |
|---|---|---|
| Rates systems running | IBM i | IBM UNIX systems with AIX |
| How IBM derives it | Measurements of a database workload held inside the IBM i Systems Performance group | An analytical model using IBM internal workloads, OLTP and SPEC benchmarks |
| Reference point | Self-referential across IBM i systems | The IBM eServer pSeries 640 is rated 1.0 |
| Disk and network I/O | The workload itself uses commitment control and journaling | The model simulates CPU, cache and memory, but not disk or network I/O |
| What IBM says it is good for | Rough model comparisons and relative CPU consumption | Approximating relative IBM UNIX commercial processing performance |
Both definitions come from the Notes on Performance Estimates page of the IBM Power11 Performance Report, 8 July 2025. The two scales are built separately for different operating systems, so one does not convert into the other.
Related Terms
IBM Power S1122
The Power S1122 is the two-socket 2U IBM Power 11 server, machine type 9824-22A, and IBM lists it for 8 or 12 Spyre AI accelerator cards ... half the ceiling of the larger S1124. For an IBM i shop it is the smaller Power 11 route into on-site model inference, with one trap: only the ERGR processors let an IBM i partition grow past 4 cores.
Dynamic Logical Partitioning
Dynamic logical partitioning, or DLPAR, moves processors, memory and card slots in and out of a running IBM Power partition without restarting it. IBM documents it for the Spyre Accelerator, so an AI card can join or leave a Linux inference partition while the IBM i production partition keeps working.
SMT8 (Simultaneous Multithreading)
SMT8 is IBM's name for running eight hardware threads at once on one Power processor core. For an IBM i shop it sets how much work an inference partition can hold, and IBM i controls it with the QPRCMLTTSK system value and the QWCCHGPR API.
AI Workload Placement
The decision about which system, operating environment, and partition executes an AI model or agent, separately from the IBM i application supplying its data.
AI Inference Partition
A separate supported operating-system partition used to execute an AI model while an IBM i partition may remain the production application and data source.
TOPS
Trillions of operations per second. A rough headline rating for AI hardware, useful for comparison and misleading on its own.