Same Cores, Same Memory, One Third of the I/O

Written and maintained by Haink's infrastructure team · Verified against Lenovo Press SR630 V4 and SR650 V4 product guides, 29 August 2026 · authorized-channel, MTM and serial verified

The 1U-or-2U question usually gets answered with a rule of thumb about rack space, and the rule of thumb is wrong in both directions. On the current generation these two servers are the same machine on the dimensions people worry about — processors, cores, memory slots, memory ceiling, memory speed — and they diverge sharply on the dimension nobody checks until detailed design.

Here is the whole decision, laid out.

Where they are identical

If your workload is bounded by cores or by memory, the chassis choice is not a performance decision. It is a decision about everything else.

Where they diverge

Three PCIe slots is the number that decides most designs. A node carrying a storage fabric, a separate management network and anything else at the same time runs out of 1U before it runs out of anything else — and the discovery usually happens during detailed design, after the platform is chosen and the price is agreed.

The density argument, run properly

"1U is denser" is true per node and misleading per rack. Compare like for like — two rack units of each:

So two rack units of 1U delivers twice the compute and twice the memory — and twice the sockets. Under per-core licensing that is the whole argument, because the licence follows the silicon. Two 1U nodes fully populated carry roughly double the annual licence of one 2U node, for the same two rack units. On a licensed virtualization estate that recurring cost usually settles the question in favour of 2U long before the slot count does. On scale-out workloads that are not licensed per core, the calculation inverts and 1U density is genuinely free.

The second thing that doubles is power. Two nodes drawing a kilowatt each is two kilowatts in two rack units, against roughly one for the 2U alternative — and rack power is the constraint that binds first in most enterprise halls, as we work through in server refresh and consolidation.

The one place 1U wins outright on this table is failure domains. Two nodes means a node failure takes out half the capacity in that space rather than all of it, which matters most on small clusters where N+1 is expensive — see sizing an HCI cluster.

Cooling: three options on 1U, a trade-off on 2U

The SR630 V4 offers three water-cooling routes: a closed-loop Processor Neptune Air Module, a Processor Neptune Core Module, and a Compute Complex Neptune Core Module. The closed-loop option is worth knowing about, because it delivers liquid cooling to the processor without requiring facility water — which removes the usual objection in a hall that has none.

On the SR650 V4 the Compute Complex Neptune Core configuration carries a cost that is easy to miss: it halves the DIMM slots from 32 to 16. On a memory-heavy host that forces larger and more expensive modules to hit the same capacity — though notably not lower bandwidth, since eight DIMMs per socket is the balanced optimum on this platform. Liquid cooling there buys density and efficiency at the price of memory capacity, not memory performance; the population rules explain the difference.

Choosing

One procurement note that applies to both: on this generation the base warranty term is encoded in the machine type — 7DG8 against 7DG9 on the SR630 V4, 7DGC against 7DGD on the SR650 V4 — so two quotes for the same chassis can carry one year or three. The liquid-cooled variants are their own machine types again, 7DK1 and 7DK2. How to read all of it is in MTM and part numbers.

Frequently asked questions

Is the SR630 V4 slower than the SR650 V4?

No. Both take the same Intel Xeon 6700 and 6500-series processors, reach 86 cores and 172 threads, carry 32 DIMM slots and 8 TB of memory, and run memory at the same 6400 MHz at 1 DPC. If the workload is bounded by cores or memory, the chassis is not a performance decision.

What do you actually give up in 1U?

Input and output. Three rear PCIe Gen 5 slots against ten, sixteen onboard NVMe ports against thirty-six, ten 2.5-inch bays or sixteen E3.S against up to forty and thirty-two respectively, and accelerator support capped at three single-wide cards at 75 W. The PCIe slot count is what decides most designs.

Isn't 1U cheaper per rack unit?

Per rack unit it delivers more compute, but it also delivers twice the sockets. Two SR630 V4 in two rack units carry four processors against two in a single SR650 V4, so under per-core licensing the annual cost roughly doubles for the same space. On core-licensed estates that usually decides it; on scale-out workloads that are not licensed per core, the density is genuinely free.

Which is better for hyperconverged nodes?

Usually the 2U. Hyperconverged nodes are storage-dense by definition, and the SR650 V4 offers up to forty 2.5-inch bays or thirty-two E3.S with thirty-two NVMe drives running without PCIe lane oversubscription. The 1U caps at sixteen E3.S.

Can we run GPUs in the 1U?

Only small ones — up to three single-wide cards at 75 W each. Anything beyond that needs the 2U chassis, which takes up to ten single-wide or two double-wide.

Does either support liquid cooling?

Both. The SR630 V4 offers three routes including a closed-loop Processor Neptune Air Module that does not require facility water, which is useful in halls that have none. On the SR650 V4, note that the Compute Complex Neptune Core configuration halves the DIMM slots from 32 to 16.

Do the two chassis have different machine types?

Yes, and more than one each. The SR630 V4 uses 7DG8 and 7DG9 for one-year and three-year base warranty and 7DK1 for the water-cooled variant; the SR650 V4 uses 7DGC and 7DGD for warranty terms, 7DK2 for Neptune Core and 7DLN for SAP HANA configurations. Compare machine types, not product names.

Related

• How to choose Lenovo ThinkSystem servers — workload-to-model selection and host sizing
• HPE DL360 vs DL380 Gen12 — 1U or 2U, and why memory is not the reason
• ThinkSystem V3 vs V4 — the generation question, and what per-core licensing does to core count
• SR650 V3 vs SR665 V3 — the Intel-or-AMD fork, if the generation is still open
• Memory population rules — both chassis carry 32 slots, and how you fill them decides the bandwidth
• Four-socket and scale-up — the same density-against-expansion trade one level up
• MTM and part numbers — why the machine type matters more than the chassis name
• Sizing an HCI cluster — failure domains, resiliency overhead and the rebuild reserve
• Server refresh and consolidation — why rack power binds before rack space
• Lenovo vs Dell vs HPE — the same chassis roles across the three vendors
• Lenovo stock, lead times and pricing · stock vs BTO/CTO

Sources

• Lenovo Press — ThinkSystem SR630 V4 product guide (1U, 86 P-cores or 144 E-cores, 32 DIMM, 8 TB, 10× 2.5″ or 16× E3.S, 16 onboard NVMe ports, 3 PCIe Gen 5 slots, 2 OCP, 3 single-wide GPUs at 75 W, three Neptune variants, machine types 7DG8 / 7DG9 / 7DK1)
• Lenovo Press — ThinkSystem SR650 V4 product guide (2U, 86 cores, 32 DIMM and 16 with Neptune Core, MRDIMM to 8000 MHz, CXL 2.0, up to 40× 2.5″ or 32× E3.S, 36 onboard NVMe ports, 10 rear PCIe slots, 2 OCP to 400 GbE, machine types 7DGC / 7DGD / 7DK2 / 7DLN)
• Lenovo Press — Introducing the ThinkSystem V4 servers with Intel Xeon 6
• VMware licensing changes (per-core subscription and the 16-core minimum per CPU, behind the socket-count argument)