Memory Stopped Being the Reason to Buy 2U

The DL360 and DL380 are the two servers most enterprise estates are actually built from, and the rule of thumb most people carry — 1U for compute, 2U when you need more memory — has been wrong for two generations. On Gen12 both machines carry 32 DIMM slots and both reach 8 TB. If memory is what is pushing you to 2U, you are paying for rack units you do not need. The real 2U arguments are drives, PCIe slots, GPUs and heat, and they are strong — but they are different arguments and they apply to different workloads.

What follows is the specification difference, then what each difference actually costs or buys in a real deployment. Figures are from HPE QuickSpecs and product pages for the two Gen12 machines.

Where they differ

Four rows carry the decision. Everything above the drive bays is effectively identical.

Compute and memory are a tie — treat them as one

Same sockets, same processor options, same 32 DIMM slots across 8 channels at 2 DIMMs per channel, same 8 TB ceiling, same speed curve. A dual-socket DL360 Gen12 and a dual-socket DL380 Gen12 with matched processors and matched memory population perform the same on compute and memory work. This has been true since Gen11, which also carried 32 slots and 8 TB in the 1U chassis.

The practical consequence: for a virtualisation estate that is CPU-and-RAM bound with modest local storage — the majority of enterprise clusters — 1U is the correct default, and every 2U you specify is a rack unit and a power budget you are spending for nothing. Across forty nodes that is a rack. The memory population question is the same on both chassis and it matters more than the chassis choice: 6400 MT/s applies only at one DIMM per channel, so a decision to fill all 32 slots costs bandwidth on either machine.

Drives: the honest 2U argument

Ten SFF bays against twenty-four, or twenty EDSFF against thirty-six. This is the largest gap between the two chassis and the one that cannot be worked around after the fact — drive capacity is set by the chassis and the backplane you order, and both are factory decisions.

Where it decides the answer:

• Software-defined storage and hyperconverged nodes. Capacity per node drives the whole cluster design: fewer, denser nodes means fewer licences but a larger failure domain and longer rebuilds. That trade is worked through in HCI node sizing, and it usually lands on 2U.
• Anything holding significant local data — backup targets, media, large local databases, object storage nodes.
• Estates that expect capacity to grow in place. Empty bays are the cheapest expansion you will ever buy. Bays you did not order are a chassis swap.

Where it does not: a virtualisation host booting from a pair of mirrored devices with all data on a SAN or an external array needs two bays and will never use the other twenty-two.

PCIe slots and GPUs: the constraint people discover late

Three rear slots plus two OCP against up to eight PCIe Gen5 slots is a wider gap than it looks, because 1U slots are also physically constrained — low-profile, and the riser configuration determines which are available at all. Count what you actually need to seat before choosing the chassis: dual-port high-speed NICs, a storage HBA or RAID controller, fibre channel, and any accelerator all compete for the same slots.

GPUs settle it outright. The 1U chassis takes single-wide, low-profile accelerators of the NVIDIA L4 class. Double-wide cards need the 2U. If there is any prospect of adding an accelerator inside the machine's life, that is a 2U decision made at purchase, not later.

One caution on 1U riser configurations: which slots exist and at what width depends on the riser you order, and risers are configured at the factory. A quote that does not state the riser configuration is not a complete quote — the same class of omission covered in sourcing HPE and part numbers decoded, where factory-integrated options cannot be added afterwards.

Heat, power and what 1U density actually costs

A 1U chassis has less air volume and shorter heatsinks, so high-TDP processors are harder to cool in it. HPE now offers closed-loop liquid cooling and direct liquid cooling as factory kits on the DL360 Gen12 — a useful capability, and also a signal about what the top of the Xeon 6 range asks of a 1U enclosure. Before finalising a dense 1U build with high-TDP parts, get two things confirmed in writing: which processor SKUs are supported with air cooling at your data centre's ambient temperature, and whether your intended configuration requires a liquid option.

Rack-level arithmetic matters as much as the chassis:

• Power and cooling per rack, not per U. Many facilities run out of kilowatts per rack long before they run out of rack units. If your limit is power, filling a rack with 1U servers is not an option regardless of what fits.
• Airflow and hot spots. Forty 1U servers in one rack is a thermal design decision, not just a mounting decision.
• Service access. 2U chassis are easier to work in, which matters more at remote sites and in colocation than in a staffed data centre.

Choosing, in order

1. Do you need double-wide GPUs, now or within the machine's life? Yes → 2U. This is absolute.
2. How many drives does the node hold? More than ten SFF or twenty EDSFF → 2U. Two boot devices and nothing else → 1U is fine.
3. Count the cards. NICs, HBAs, fibre channel, accelerators. If the total exceeds what the 1U riser configuration supports, that is a 2U.
4. Check the rack's power ceiling. If power per rack is the binding constraint, 1U density buys you nothing and 2U costs you nothing.
5. Confirm thermals for your chosen processors in 1U at your site's ambient, in writing.
6. Otherwise take the 1U. Same compute, same memory, half the rack.

One note on mixed fleets: many estates end up standardising on 2U for everything to keep spares, images and procedures uniform. That is a defensible operational choice, not a technical one, and it is worth pricing the rack space it costs before adopting it by default — particularly on multi-node orders, where the difference compounds. The sequence for orders at that scale is in buying 20–60 servers as one project.

Frequently asked questions

Does the DL380 hold more memory than the DL360?

No. Both Gen12 machines carry 32 DIMM slots — 8 channels per socket at 2 DIMMs per channel — and both reach 8 TB, 4 TB per processor. This was already true on Gen11. If memory capacity is the reason you are specifying 2U, the reason does not hold.

Is the DL360 Gen12 dual-socket?

Yes. It supports two Intel Xeon 6 processors, with the second optional, in a 1U chassis. Single-socket is a configuration choice on both machines, not a chassis limitation of the 1U.

Can I add a GPU to a DL360 later?

Only a single-wide, low-profile accelerator of the NVIDIA L4 class, and only if a suitable slot is free in the riser configuration you ordered. Double-wide cards do not fit a 1U chassis at all. If an accelerator is a possibility within the server's life, specify 2U at purchase — the alternative is replacing the machine.

How many drives does each take?

DL360 Gen12: up to 10 × 2.5″ SFF, or 4 LFF, or up to 20 EDSFF E3.S NVMe. DL380 Gen12: up to 24 SFF front, or 12 LFF, or up to 36 EDSFF E3.S. Drive capacity is fixed by the chassis and backplane at order time and cannot be raised later.

Which is better for VMware or another virtualisation cluster?

Usually 1U, if data lives on shared storage. Compute and memory are identical, so the 2U buys you drive bays a SAN-backed host will not use. The exception is hyperconverged or software-defined storage, where node capacity drives the cluster design and 2U is normally correct. Note that under per-core licensing the processor variant matters far more to your total cost than the chassis — see hardware selection after the VMware licensing change.

Does 1U cost less than 2U?

The chassis difference is modest and is not where the money is. The real economics are rack space, power and cooling over the machine's life, and they favour 1U only if your facility is constrained by rack units rather than by kilowatts per rack. Check which constraint is actually binding before optimising for the other one.

Can I mix DL360 and DL380 in the same cluster?

Yes — same generation, same processors, same memory architecture, so a mixed cluster behaves uniformly provided the processor SKUs and memory populations match. This is a common and sensible pattern: 1U for compute-only nodes, 2U where a node needs local capacity or an accelerator. Keep firmware baselines aligned across both.

Related

• HPE ProLiant Gen11 vs Gen12 — the core-variant and memory-speed decision that comes before the chassis
• 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
• HPE memory population rules — why 12 DIMMs per processor runs at half the bandwidth of 8
• HPE Alletra explained — what happened to 3PAR, Primera and Nimble, and the lifecycle dates
• How to choose HPE ProLiant servers and storage — platform selection and host sizing
• HPE part numbers decoded · sourcing HPE · GreenLake vs owning
• HCI node sizing · after the VMware licensing change · refresh and consolidation
• The same 1U/2U decision on Lenovo — SR630 V4 vs SR650 V4
• Dell vs HPE · Lenovo vs Dell vs HPE
• Buying 20–60 servers as one project · stock vs BTO/CTO
• HPE stock, lead times and pricing — current availability and firm quotes

Sources

• HPE — ProLiant Compute DL380 Gen12 (up to 144 cores; up to 8 TB DDR5 at up to 6400 MT/s; up to 36 EDSFF E3.S; up to three double-wide or eight single-wide GPUs; direct liquid cooling)
• HPE — ProLiant Compute DL360 Gen12 QuickSpecs (1U; two sockets; "32 DIMMs in total", "8 channels DDR5 per socket", "2 DIMMs-per-channel", "8TB max memory limit (4TB per processor)", "Up to 6400MTs 1DPC, 5200 2DPC"; 10 SFF or 4 LFF or 20 E3.S; closed-loop and direct liquid cooling kits; iLO 7)
• HPE — ProLiant DL360 Gen11 QuickSpecs (32 DIMM slots and 8.0 TB in the 1U chassis on the previous generation)