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.
What does SMT8 do?
Simultaneous multithreading, usually shortened to SMT, lets one processor core work on more than one instruction stream at the same time. IBM i documentation describes it as sharing processor facilities so that two applications, or two threads of the same application, run together. SMT8 is the eight thread setting on Power 10 and Power 11 cores. SMT4 is four threads, SMT2 is two, and single thread mode is one.
Think of a core as one kitchen. In single thread mode, one cook has the whole kitchen. In SMT8, eight cooks share the same stove and the same counters. Many more orders move through that kitchen in an hour. No single order comes out any faster.
IBM says this plainly in its own IBM i documentation: running multiple instruction streams at the same time does not improve the performance of any given task. SMT raises how much total work a partition can hold, not the speed of one job.
How much more work fits in one partition?
What IBM rates a 16 core Power 11 S1124 at, by SMT setting
IBM Power Performance Report, Power7 to Power11, dated July 14, 2026, for machine type model 9824-42A with 16 Power 11 cores at 3.40 to 4.20 GHz. rPerf is IBM's own estimate of commercial processing performance for its UNIX systems, with the eServer pSeries 640 set at 1.0. IBM states it is not a public benchmark result. IBM i capacity planning uses CPW instead.
Single thread against SMT8 across the Power 11 S1124 line
| Cores in the server | Single thread | SMT8 | Rated gain |
|---|---|---|---|
| 16 cores at 3.40 to 4.20 GHz | 156.8 rPerf | 572.9 rPerf | 3.7 times |
| 32 cores at 3.40 to 4.20 GHz | 305.8 rPerf | 1,117.1 rPerf | 3.7 times |
| 48 cores at 3.05 to 4.15 GHz | 435.6 rPerf | 1,529.0 rPerf | 3.5 times |
| 60 cores at 2.80 to 3.95 GHz | 515.8 rPerf | 1,736.8 rPerf | 3.4 times |
Same IBM source. The rated gain column divides IBM's SMT8 figure by its single thread figure for the same machine. Eight times the threads is not eight times the work, because those threads share one core.
How do you set SMT on IBM i?
Three controls, in the order you touch them
- QPRCMLTTSK, the on and off switchThis IBM i system value has three settings. Off means each processor handles only one task at a time. On means processors handle several tasks at once. System controlled, the default, also lets the operating system occasionally hold a processor to a single task. On Power 6 hardware and later, the change takes effect immediately.
- QWCCHGPR, the thread ceilingThis API sets the highest number of threads a core may use. IBM lists the values: 7 for SMT8, 3 for SMT4, and 1 for SMT2. IBM states the change is effective immediately and persists across a partition restart.
- QWCRRTVPR, the checkThis API reads back the maximum number of secondary threads, so you can confirm the partition really is in the mode you set.
Values and behavior from IBM's own SMT2, SMT4 and SMT8 support page for IBM i, which also records that IBM i 7.1 defaulted to SMT4 and IBM i 7.2 moved the default to SMT8. The QPRCMLTTSK settings and the Power 6 note come from IBM i documentation for the processor multitasking system value.
Does SMT8 make AI inference faster?
Not on its own. When a model runs on the cores, the heavy math goes through Matrix Math Acceleration, and IBM's own Power 10 briefing material counts four matrix engines in each core. Every thread on that core shares those same four engines. Switching on more threads gives the partition more places to queue work. It does not add engines.
When a Spyre Accelerator card is doing the inference, the math leaves the core and runs on the card. The SMT setting then shapes the partition around the card, such as how many requests the serving process can keep in flight, rather than the speed of the card itself.
- More threads per core means more total work in flight, not a faster answer to one request
- The four matrix engines inside a core are shared by every thread running on that core
- IBM's Power 11 performance checklist, last updated 16 March 2026, tells shops moving to Power 11 to consider SMT8 and to size partitions on the SMT8 rPerf values, which it says will often reduce the virtual processors a partition needs
IBM also warns that SMT gains depend heavily on the application and the model, that results vary between environments, and that some applications are better served with the system value switched off. Measure your own workload before you settle on a setting.
Related Terms
Matrix Math Acceleration (MMA)
An on-chip acceleration feature in IBM Power10 and Power11 processors that speeds up the matrix math operations used in AI inferencing, without requiring a separate GPU.
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.
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.
IBM Power S1124
The Power S1124 is the largest scale-out IBM Power 11 server, machine type 9824-42A, and IBM lists it for 8, 12, 16 or 24 Spyre AI accelerator cards ... but only when both processor modules are installed. An IBM i shop can keep production on IBM i 7.4 TR12 or later and run model inference in a Red Hat Enterprise Linux partition on the same 4U machine.