Twelve DIMMs Per Processor Runs at Half the Bandwidth of Eight

On HPE Gen12 servers with Intel Xeon 6, HPE's own population table puts eight DIMMs per processor at 100% of peak memory throughput and twelve DIMMs at 50% — the same figure it gives for four. Adding four modules to a correctly populated server does not add a little performance or none. It halves the memory bandwidth while increasing the bill. This is the most expensive configuration mistake we see on server quotes, and it never appears as a warning on a bill of materials.

What follows is the population arithmetic for both processor families, the mixing rules that will get an order rejected or a server refusing to boot, and how to write the memory line of a quote so that two suppliers are actually quoting the same machine. Figures are from HPE's technical white papers on server memory population and from the model QuickSpecs.

The Intel Xeon 6 table

Gen12 servers on Xeon 6 provide 8 memory channels per processor and 16 DIMM slots per processor at 2 DIMMs per channel. HPE publishes the supported populations and what each one costs you:

HPE's position is unambiguous: "Hewlett Packard Enterprise discourages unbalanced configurations because they will always have lower performance than similar balanced configurations."

Why twelve is worse than eight

Memory bandwidth is a property of channels, not of gigabytes. With 8 channels per processor, the controller interleaves across all channels that are populated identically. Eight DIMMs means one per channel — every channel working, perfectly symmetric, full bandwidth.

Twelve DIMMs cannot divide evenly across eight channels. Four channels end up with two modules and four with one, so the interleaving pattern breaks and the memory controller falls back to the pace the asymmetric region can sustain. You have paid for 50% more memory and received half the bandwidth of the eight-DIMM machine. The capacity is real; the throughput is not.

This is also why the intuition "more sticks means more speed" fails here and holds on desktops: a desktop has two channels and you are usually going from one populated to two, which is the move from unbalanced to balanced. On a server with eight channels, the equivalent move is from four DIMMs to eight, and after that every step that is not a doubling makes things worse.

The two configurations that make sense

For practical purposes, a Xeon 6 processor should hold either 8 DIMMs or 16. Everything else is a compromise you should be choosing deliberately, not falling into.

Note what that means against the twelve-DIMM case: sixteen DIMMs is balanced and lands within roughly a fifth of peak, while twelve is unbalanced and halves it. Filling every slot is a reasonable trade; stopping four short of it is not.

The commercial trap is that 16 smaller DIMMs usually price lower than 8 larger ones for the same total capacity, so a supplier optimising the quote total will reach for 2 DPC by default. That is a legitimate configuration — it is balanced — but it is a performance decision being made inside a pricing decision, and it should be yours. The generation-level context is in Gen11 versus Gen12, where the same 1 DPC condition governs whether the newer platform is faster at all.

One further note on 8 versus 16: leaving 16 slots empty is the cheapest expansion path in the server. Doubling memory later costs modules; going from a full 16 to more capacity costs a forklift.

AMD is a different shape entirely

The AMD Gen12 servers — DL325 Gen12 and DL345 Gen12, both single socket — run 5th Gen EPYC with 12 memory channels per processor and 24 DIMM slots. The balanced point is therefore 12 DIMMs, not 8, and HPE quotes up to 6 TB of DDR5 at 5200 MT/s on the DL345 Gen12.

Two consequences for anyone cross-shopping:

• A balanced AMD socket needs 12 modules; a balanced Intel socket needs 8. Comparing "a server with 12 DIMMs" across the two families compares a correct configuration against a broken one. This is a real source of misleading quotes.
• One AMD socket carries 50% more channels than one Intel socket. For bandwidth-bound single-socket work that is a genuine architectural advantage, and it pairs with the licensing argument for single-socket machines covered in Gen11 vs Gen12.

HPE publishes the AMD population rules separately, in the technical white paper for ProLiant Gen11 and Gen12 with EPYC 9005 series processors — document a50012817enw. Ask for the current revision rather than assuming the Intel table transfers, because it does not.

Mixing rules that stop an order

These are not guidelines. Violating them produces a system that will not boot, will not train memory reliably, or will not be supported:

• RDIMM mixes only with RDIMM. Do not mix RDIMMs and LRDIMMs in the same system.
• MRDIMM mixes only with MRDIMM. Do not mix MRDIMMs with RDIMMs or LRDIMMs.
• Do not mix x4 with x8 within a channel or across channels.
• Do not mix die densities in the same platform — 24 Gb die with non-24 Gb die, or 32 Gb die with non-32 Gb die.
• Do not mix non-3DS and 3DS in the same platform.
• Rank mixing on a channel is not allowed except where every DIMM slot on each channel is fully populated. Some Xeon 6 parts prohibit rank mixing outright.

The practical consequence appears at upgrade time, not at purchase. Memory bought two years ago from a different SKU line frequently cannot join the modules already in the machine, even at the same capacity, because the die density or organisation differs. If you plan to expand later, record the exact part numbers now — the option-versus-spare distinction and the -B21 against -F21 memory SKU split are covered in HPE part numbers decoded.

Population order

Within each channel, the white connector slot is populated first, then the black. Where a server has more than one processor, split the DIMMs evenly between them and apply the same rule to each.

Uneven memory across processors causes a subtler problem than uneven memory across channels. HPE describes it directly: threads running on the processor with less memory may consume everything local to it and then request remote memory from the other processor, producing longer latencies and limited throughput. In a virtualisation cluster this shows up as unexplained latency on some virtual machines and not others, and it is invisible in capacity monitoring because the total looks fine.

Specifying memory so quotes are comparable

A memory line reading "512 GB" is not a specification. Two servers with 512 GB can differ by a factor of two in bandwidth and by a full generation in upgrade path. State these five things and the quotes become comparable:

1. Total capacity per server, and whether it must be reached without using the second DIMM per channel.
2. Module count and size — "16 × 32 GB" and "8 × 64 GB" are different machines.
3. Type and organisation — RDIMM or MRDIMM, x4 or x8, and whether 3DS.
4. Whether free slots are required for a planned expansion, and how much headroom.
5. Whether the workload is bandwidth-bound or capacity-bound. If you tell a supplier only the capacity, you have delegated this decision to them.

If a quote gives capacity without module count, it is not answerable. Ask for the population before comparing totals — one of the checks in what a complete HPE quote contains.

Frequently asked questions

How many DIMMs should a Gen12 server have per processor?

Eight or sixteen on Intel Xeon 6, twelve or twenty-four on AMD EPYC 9005. Eight per Intel processor is the balanced 1 DPC configuration at 100% of peak throughput and the highest clock; sixteen fills every slot and is also balanced, at a lower clock. Counts between those, particularly twelve on Intel, are unbalanced and carry a real performance penalty.

Why is 12 DIMMs per processor bad on Intel?

Because the processor has 8 memory channels and twelve modules cannot spread evenly across them. Four channels get two DIMMs and four get one, the interleave breaks, and HPE's own throughput column places the result at 50% of peak — the same as a four-DIMM machine. You pay for 50% more memory and lose half the bandwidth.

Can I mix DIMM sizes in the same server?

Only within tight limits, and usually you should not. RDIMM mixes only with RDIMM and MRDIMM only with MRDIMM; x4 and x8 cannot be mixed within or across channels; die densities cannot be mixed in the same platform; and rank mixing on a channel is only permitted when every slot on each channel is populated. In practice a mixed-capacity configuration is also unbalanced, so even where it is legal it is slow.

Does adding memory later work?

It works if you plan for it at purchase: populate 1 DIMM per channel, leave the second slot per channel empty, and record the exact module part numbers. It fails when the original modules are no longer available in the same die density and organisation, which is common two or three years on. Buy the expansion modules with the server if the expansion is certain.

Is 2 DPC always slower than 1 DPC?

Per channel, yes — the clock drops. On the DL380 Gen12 the 8 TB configuration is specified at 5200 MT/s against 6400 MT/s at 1 DPC. But 2 DPC is balanced and uses every channel, so it is the right answer when total capacity is the binding constraint. It is the wrong answer when the workload is bandwidth-bound and the capacity would fit in 1 DPC using larger modules.

Do these rules apply to Gen11 as well?

The principle does — balance across channels, no mixing across types and densities, white slots first — but the specific tables differ by processor family and generation, and HPE publishes a separate paper for each. Do not carry a Gen12 Xeon 6 table onto a Gen11 4th or 5th Gen Xeon server; request the paper matching your exact platform.

Does memory population affect software licensing?

Not directly — hypervisors and databases license by core or socket, not by DIMM. It matters indirectly and substantially: if a bandwidth-starved configuration forces you into more hosts to hit the same performance, you pay for those extra cores. Sizing memory correctly is one of the cheapest ways to keep the host count down, which is where per-core licensing bites — see hardware selection after the VMware licensing change.

Related

• HPE ProLiant Gen11 vs Gen12 — why the memory-speed gain only exists at 1 DPC
• HPE four-socket scale-up — DL580 Gen12 against two DL380, and why 4U buys only memory
• HPE support tiers explained — why four-hour response is weaker than six-hour repair
• HPE iLO and Compute Ops Management licensing — the remote console is a per-server licence
• DL360 vs DL380 Gen12 — both chassis carry the same 32 slots, so this decision is chassis-independent
• HPE part numbers decoded · sourcing HPE · choosing HPE platforms
• The same arithmetic on Lenovo ThinkSystem — identical Intel channel counts, different slot maps
• After the VMware licensing change · refresh and consolidation · HCI node sizing
• Buying 20–60 servers as one project · stock vs BTO/CTO
• HPE stock, lead times and pricing — current availability and firm quotes

Sources

• HPE — Server memory population rules for HPE Gen12 servers with Intel Xeon 6 processors (8 channels and 16 DIMM slots per processor; the supported population table and its throughput column — 1 DIMM 12.5%, 2 DIMMs 25%, 4 DIMMs 50%, 8 DIMMs 100%, 12 DIMMs 50%; white connector populated first; the RDIMM/MRDIMM, x4/x8, die-density, 3DS and rank-mixing rules; "Hewlett Packard Enterprise discourages unbalanced configurations because they will always have lower performance than similar balanced configurations"; the cross-processor remote-memory latency note)
• HPE — Server memory population rules for HPE ProLiant Gen11 and Gen12 servers with AMD EPYC 9005 series processors (the AMD population tables; request the current revision)
• HPE — ProLiant Compute DL345 Gen12 (24 DIMM slots, up to 6 TB DDR5 at 5200 MT/s, single socket, 12 channels per processor)
• HPE — ProLiant Compute DL380 Gen12 (up to 8 TB DDR5 at up to 6400 MT/s; 5200 MT/s at 2 DPC in the QuickSpecs maximum-capacity configuration)
• Lenovo Press — Balanced memory configurations for 2-socket servers with Intel Xeon 6 (the 75–82% figure for a fully populated 2 DPC configuration; cited because it measures the same Intel memory architecture and HPE does not publish an equivalent percentage)