Differential DIMM (DDIMM)
A DDIMM is the memory card an IBM Power 11 or Power 10 server takes in place of a plain DDR memory stick. It carries its own buffer chip and plugs into an Open Memory Interface slot, which is how a two-socket Power S1124 reaches the 2456 GBps memory bandwidth ceiling IBM lists for feeding data into an AI model.
What is a DDIMM?
A plain memory stick is just memory chips on a card, and the processor has to talk to every chip on it. A DDIMM puts a buffer chip on the card as well, so the processor talks to the buffer and the card handles the chips behind it. Think of a hotel front desk. You ask one person at the desk instead of knocking on 32 doors. IBM describes DDIMMs as high-performance, high-reliability, high-function memory cards that contain a buffer chip, intelligence, and 4000 MHz or 4800 MHz DRAM memory.
- DDIMM stands for differential DIMM. DIMM is dual inline memory module, the industry name for a memory stick.
- The card plugs into an Open Memory Interface slot, not a standard DDR slot. IBM also writes OMI as the OpenCAPI Memory Interface.
- OMI is technology-neutral, so the interface is not tied to one memory generation. The same design carries DDR4 or DDR5 chips.
- IBM groups OMI and DDIMMs with Transparent Memory Encryption on Power 11 scale-out servers, so memory performance, resilience and security are sold as one story.
Why does a memory card matter for AI?
An AI model does not sit still while it answers. Every answer means pulling model weights through the processor. If memory cannot hand data over fast enough, the matrix engines sit and wait. IBM says this plainly: the on-chip Matrix Math Acceleration engines and vector instructions work combined with high memory bandwidth, which IBM calls a critical component for feeding data into AI models rapidly.
IBM puts memory bandwidth behind its biggest AI claim for Power 11. SG24-8590 says Power 11 holds a 10x to 20x performance advantage over Power 9 for AI inferencing workloads, driven by increased memory bandwidth and new instructions. The same book lists up to a 50 percent increase in memory bandwidth from DDR5 and advanced memory controllers, and twice the OMI memory bandwidth of the Power 10 platform. These are IBM figures against IBM's own older machines, not a measured result on your workload.
How much bandwidth does each DDIMM size give?
Bigger is not always faster. On a Power S1124 the 64 GB DDIMM is the card that reaches IBM's top number. Step up to a 128 GB or 256 GB card and the listed channel rate drops from 4800 to 4000 Mbps, so the ceiling comes down with it. You buy more room and give back speed.
Maximum theoretical memory bandwidth per system on a Power S1124
All four figures come from one comparison table in IBM Redbooks publication SG24-8590, which lists the same 4U Power S1124 chassis with two Power 10 dual-chip modules and with two Power 11 dual-chip modules. Mbps here is the data rate IBM lists per OMI channel, and a two-socket system has 32 channels. One caution: the same book's Power S1124 section says the slots deliver up to 1228 GBps, which does not match its own table. Confirm the figure for the exact memory feature code you are quoting.
Can you reuse memory you already own?
Sometimes, and that is the whole point of OMI. IBM says Power 11 servers also support OMI DDR4 memory migrated from Power 10 high-end systems, because OMI is technology-neutral, and it gives the reason as protecting client investments in memory technology. The feature codes are not the same between generations even where the slot is, so ask your reseller with your machine type in hand instead of assuming the cards move.
How do you order DDIMMs?
DDR5 memory features IBM lists for the Power S1122 and Power S1124
| Feature code | What you get | Memory chip size |
|---|---|---|
| EM54 | 64 GB as two 32 GB DDIMMs | 16 Gbit DDR5 |
| EM5B | 128 GB as two 64 GB DDIMMs | 16 Gbit DDR5 |
| EM5G | 256 GB as two 128 GB DDIMMs | 32 Gbit DDR5 |
From the system feature code tables in IBM Redbooks publication SG24-8590, which lists all three at 4000 MHz or 4800 MHz. Every one of these feature codes delivers two physical memory cards, which is why the rules below count in pairs and quads. The high-end Power E1180 carries its own memory feature codes and is not covered by this table.
Four rules IBM sets before the memory will go in
- One processor module: 32 GB minimum, ordered in pairsIBM sets a floor of 32 GB of memory with one processor module installed, and says all memory DIMMs must be ordered in pairs.
- Two processor modules: 64 GB minimum, ordered in quadsWith both sockets filled the floor doubles to 64 GB and the cards must be ordered in quads. A quad is two feature codes, because one feature code is already two cards.
- Fill every slot if you care about bandwidthIBM's own guidance is to populate all DIMM slots for the best possible performance, especially for OLAP and other high-bandwidth workloads. OLAP is online analytical processing, the reporting and analysis side of a database. Model inference reads memory with the same appetite.
- Check the system firmware firstIBM notes that DDR5 memory DDIMMs require firmware FW1110.00 or later. That is a check to make before the cards arrive, not after.
There is one more trap on the smaller processor option. IBM says only input and output is supported on chip 0 of an entry single-chip module, so an eSCM socket takes eight DDIMMs, for a maximum of 16 per system, against up to 32 on a dual-chip module. Fewer slots means fewer OMI channels, and the bandwidth ceiling falls with them.
Related Terms
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.
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.
IBM Power E1180
The Power E1180 is the high-end IBM Power 11 server, machine type 9080-HEU, and the only Power model IBM lists for more than 24 Spyre AI accelerator cards ... up to 48 per system. An IBM i shop can keep production on IBM i 7.4 TR12 or later and run large AI inference in a Red Hat Enterprise Linux partition on the same machine.
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 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.