Filling Every Slot Costs You a Fifth of Your Memory Bandwidth. On the Other Platform It Gains You a Third.

Written and maintained by Haink's infrastructure team · Bandwidth figures from Lenovo Press balanced-memory papers for Intel Xeon 6 and 5th Gen AMD EPYC, 29 August 2026 · authorized-channel, MTM and serial verified

"Add more memory" reads like a capacity decision. It is also, silently, a bandwidth decision — and the correct answer is opposite on the two platforms Lenovo sells in the same chassis.

On Intel, peak memory bandwidth arrives at half the slots populated, and filling the rest costs you performance. On AMD, peak arrives only when every slot is populated. Both figures come from Lenovo's own balanced-memory papers, and the number of people specifying servers who know this is small.

Intel Xeon 6: peak is at 8 DIMMs per socket, not 16

Each Xeon 6 processor has eight memory channels and up to two DIMM slots per channel — sixteen slots per socket, thirty-two on a two-socket server. Relative memory bandwidth by population:

Two rows deserve attention. Sixteen DIMMs per socket — every slot filled — runs at 75–82% of the bandwidth that eight DIMMs deliver. Doubling capacity costs roughly a fifth of the bandwidth, because the second DIMM per channel slows the channel down.

And twelve DIMMs per socket is unbalanced and can drop to 50% — worse than eight, for half again as much memory. Twelve is exactly where you land by taking a working eight-DIMM configuration and adding four more sticks, which is the single most common way a server gets slower after an upgrade that was supposed to make it faster.

The three rules behind the table

Lenovo states them plainly, and they are worth quoting because they explain every number above:

1. "All populated memory channels should have the same total memory capacity and the same number of ranks per channel"
2. "All memory controllers on a processor socket should have the same configuration of memory DIMMs"
3. "All processor sockets on the same physical server should have the same configuration of memory DIMMs"

Which is why mixed capacities, odd counts and "we had some spare DIMMs" configurations underperform in ways that look like a mystery until someone counts the sticks.

5th Gen AMD EPYC: peak is at 12, and every step down costs you

The AMD platform is built differently. Twelve memory channels per processor, and one DIMM per channel on Lenovo ThinkSystem servers — so twelve slots per socket, twenty-four on a two-socket server, with no second DIMM per channel available at all.

Bandwidth here tracks channel count almost linearly. There is no clever half-populated optimum — every channel you leave empty is bandwidth you did not buy.

The number twelve means opposite things

This is the trap worth remembering, because it survives across every conversation about these two platforms:

Same number of sticks, same chassis size, opposite verdict. A memory configuration carried across from an AMD design to an Intel one, or the reverse, is not a translation — it is a new calculation.

How to specify the memory, in order

1. Start from the balanced count, not the capacity. Eight DIMMs per socket on Intel, twelve on AMD. That fixes the number of modules; capacity then comes from module size.
2. Reach the capacity target with module size. Need 1 TB per socket on Intel? That is eight 128 GB modules, not sixteen 64 GB ones — same capacity, better bandwidth, and usually a similar price once the larger module premium is weighed against buying twice as many.
3. Only go to sixteen per socket on Intel when capacity genuinely demands it, and price the 75–82% bandwidth into the decision. For memory-bound workloads that trade is frequently wrong.
4. Never land on twelve per socket on Intel. If the capacity maths points there, go to sixteen with smaller modules or eight with larger ones — both beat it.
5. Keep every socket identical. Rule three above is not advisory; asymmetric sockets underperform on both platforms.
6. Confirm the rules for MRDIMM separately. The SR650 V4 supports MRDIMMs running to 8000 MHz, and population rules for them should be checked against the guide for the specific platform rather than assumed to match RDIMM behaviour.

Planning for expansion, which works differently on each

On Intel there is genuine headroom: specify eight per socket now, and sixteen is available later. The cost of that expansion is the bandwidth drop, not a hardware swap — so it is a legitimate strategy for an estate whose capacity needs are uncertain.

On AMD, with one DIMM per channel and twelve slots, a fully balanced node is already fully populated. Adding capacity later means replacing modules with larger ones, and the removed modules have no home. Which makes the module-size decision on an AMD node considerably more consequential at purchase than it is on an Intel one — a point worth carrying into the Intel-or-AMD comparison, where the AMD platform's lower memory ceiling is already part of the trade.

The liquid cooling footnote, corrected

On the SR650 V4, the Compute Complex Neptune Core configuration reduces the server from 32 DIMM slots to 16. That is usually described as a straight loss, and on capacity it is — the ceiling comes down.

On bandwidth it is not. Sixteen slots on a two-socket machine is eight per socket, which is precisely the balanced, 100%-bandwidth configuration. So the liquid-cooled variant costs you the ability to reach the highest capacities and costs you nothing in memory performance, provided you populate it correctly. That is a materially better trade than "half the slots" makes it sound, and it changes the calculation for anyone choosing liquid cooling on a workload that is bandwidth-sensitive rather than capacity-hungry.

What to check on a quote

• DIMM count per socket, stated explicitly — not total capacity.
• Module capacity and rank count, identical across all populated channels.
• Both sockets configured identically.
• The rated speed at the quoted population — 1 DIMM per channel and 2 DIMMs per channel run at different speeds on the Intel platform, and the difference belongs on the quote.
• Whether the configuration is factory or field-populated, because a machine arriving correctly populated is one fewer thing done forty times later.

Frequently asked questions

Should we fill every memory slot?

On AMD, yes — twelve DIMMs per socket is the balanced optimum and delivers 100% of available bandwidth. On Intel, no: eight per socket delivers 100%, and filling all sixteen drops to 75–82%. Fill all sixteen only when capacity genuinely requires it.

What is the worst memory configuration on an Intel server?

Aside from a single DIMM, which runs at 13–15%, the practical worst case is twelve DIMMs per socket — unbalanced, and as low as 50%. It is also the configuration people land on by adding four sticks to a working eight-DIMM server. Go to sixteen or stay at eight.

How do we reach a large capacity without losing bandwidth on Intel?

With larger modules rather than more of them. Eight 128 GB DIMMs per socket beats sixteen 64 GB DIMMs for the same 1 TB, because the eight-DIMM population runs at full bandwidth while the sixteen-DIMM one runs at 75–82%.

Why is twelve DIMMs good on AMD and bad on Intel?

Channel count. 5th Gen EPYC has twelve memory channels per processor, so twelve DIMMs means one per channel — perfectly balanced. Xeon 6 has eight channels with up to two DIMMs each, so twelve DIMMs loads some channels twice and others once. Same number, opposite meaning.

Can we add memory to an AMD node later?

Only by replacing modules. Lenovo ThinkSystem AMD platforms support one DIMM per channel with twelve slots per socket, so a balanced node is already fully populated. The removed modules have no home, which makes module size a more consequential decision at purchase than it is on Intel.

Does liquid cooling hurt memory performance on the SR650 V4?

No. The Compute Complex Neptune Core configuration reduces the server to sixteen DIMM slots, which on a two-socket machine is eight per socket — exactly the balanced, full-bandwidth population. It costs capacity headroom, not bandwidth.

What happens if the two sockets are populated differently?

Performance drops. Lenovo's balancing rules require all processor sockets on the same server to have the same DIMM configuration, alongside equal capacity and rank count on every populated channel. Asymmetric sockets are one of the more common causes of memory performance that nobody can explain.

Related

• SR650 V3 vs SR665 V3 — where the 32-slot and 24-slot difference between the platforms actually matters
• HPE memory population rules — why 12 DIMMs per processor runs at half the bandwidth of 8
• SR630 V4 vs SR650 V4 — both carry 32 slots, so memory is not the chassis differentiator
• ThinkSystem V3 vs V4 — memory speed by generation, and why the ceiling did not move
• Four-socket and scale-up — 64 DIMM slots and a 16 TB ceiling, when one instance needs more than 8 TB
• Sizing an HCI cluster — why memory usually binds before CPU
• How to choose Lenovo ThinkSystem servers · MTM and part numbers
• Lenovo stock, lead times and pricing · memory and components in stock

Sources

• Lenovo Press — Balanced memory configurations for 2-socket servers with Intel Xeon 6 (eight channels per processor, two DIMMs per channel; the three balancing rules; relative bandwidth at 1, 4, 8, 12 and 16 DIMMs per socket)
• Lenovo Press — Balanced memory configurations with 5th Gen AMD EPYC (twelve channels per processor, one DIMM per channel; populated channel counts of 1, 2, 4, 6, 8, 10 or 12; relative bandwidth per DIMM count at NPS1; unbalanced as low as 9%)
• Lenovo Press — Balanced memory configurations for 4th and 5th Gen Intel Xeon Scalable (the previous generation, for V3 platforms)
• Lenovo Press — ThinkSystem SR650 V4 product guide (32 DIMM slots, 16 with Compute Complex Neptune Core; 6400 MHz at 1 DPC and 5200 at 2 DPC; MRDIMM to 8000 MHz)
• Lenovo Docs — SR650 V4 independent mode installation order (the physical population sequence)