Gen12 Adds Cores You May Not Want to License

The headline on HPE ProLiant Compute Gen12 is 144 cores per socket against 64 on Gen11. Read that as a straight upgrade and you can walk into a configuration that costs more in hypervisor licensing than the servers saved you, while delivering fewer threads than the part you skipped. The generational change is real and in several places substantial — but the reason to move is rarely the core count, and the core count is where most Gen11-versus-Gen12 comparisons stop.

This page covers what changed on the two-socket rack line, which differences alter a quote, and the cases where ordering Gen11 in 2026 is still the correct decision. Figures are from HPE QuickSpecs and product pages for the DL380 Gen11 and DL380 Gen12; the reasoning applies across the DL300 line.

What actually changed

Three lines in that table do work in a real decision. The DIMM topology and the memory ceiling did not move. The NVMe density tripled. And the management processor changed in a way that matters more to security review than to operations.

The core-count line, read properly

Intel's Xeon 6 family splits into two designs sold under one name. P-core parts are the conventional performance cores with simultaneous multithreading; E-core parts are dense efficiency cores with no multithreading. The 144-core figure in HPE's specification comes from the E-core side. A P-core Xeon 6 in the same socket lands considerably lower on core count but presents two threads per core.

For a per-core-licensed platform this reverses the intuition:

• 144 E-cores gives you 144 threads and 144 licensable cores.
• A high-end P-core part in the same socket gives you fewer cores — and more threads.

Under VMware's post-Broadcom per-core subscription, Nutanix per-core licensing, or any database licensed by core, the E-core part is the more expensive machine to run despite being the cheaper machine to buy, and it delivers fewer threads to the scheduler. E-core parts earn their place in throughput-oriented, licence-light estates — web tiers, containerised microservices, scale-out infrastructure you own the software for. They are usually the wrong answer under a commercial hypervisor. The same split exists on the Lenovo side of the market and we worked the arithmetic through there: ThinkSystem V3 versus V4 covers the identical Xeon 6 decision on a different badge, and hardware selection after the VMware licensing change covers what per-core pricing does to node sizing generally.

Practical rule: decide the licence model first, then pick the core variant, then pick the generation. Doing it in the other order is how a refresh ends up with a five-figure annual licensing increase nobody forecast.

Memory: same ceiling, faster — until you fill the slots

Both generations carry 32 DIMM slots and both top out at 8 TB. If you are moving to Gen12 for capacity headroom, there is none to gain on this platform. What changed is speed, and by less than the marketing implies. Gen11 on 5th Gen Xeon already reaches up to 5600 MT/s depending on processor and population, falling to 4800 MT/s in the fully populated 8 TB configuration. Gen12 raises that to 6400 MT/s — at one DIMM per channel only. Populate the second DIMM per channel and the specification drops to 5200 MT/s, which is below what a well-configured Gen11 can do at 1 DPC.

That turns a routine configuration choice into a measurable trade:

• 16 DIMMs (1 DPC), 8 per socket — full 6400 MT/s. To reach a given capacity you buy larger, more expensive modules.
• 32 DIMMs (2 DPC) — maximum capacity and cheaper modules per gigabyte, at 5200 MT/s.

For memory-bandwidth-sensitive work — in-memory databases, analytics, dense virtualisation with high consolidation ratios — the 1 DPC route is usually correct even though the DIMMs cost more. For capacity-bound estates, 2 DPC is fine. The mistake is ordering 2 DPC by default because it looked cheaper on the bill of materials, then troubleshooting bandwidth for a year. The same population logic, with the channel arithmetic laid out, is in memory population rules.

Note also what this does to a quote: two configurations reaching the same capacity can differ substantially in price and in performance, and neither is wrong. If a supplier quotes you memory without saying which population it assumes, the quote is not comparable to any other quote — one of the checks in sourcing HPE.

iLO 6 to iLO 7

iLO 7 introduces a Secure Enclave — an independent security processor holding credentials and configuration in tamper-resistant hardware rather than in firmware alone — with FIPS 140-3 Level 3 support alongside the Silicon Root of Trust that has been present since iLO 4. Gen12 servers running iLO 7 firmware meet NIST SP 800-208, the stateful hash-based signature standard that makes firmware signing resistant to a future quantum attack. HPE also reports materially faster management operations and boot.

Whether this justifies a generation change depends entirely on who reviews your architecture. If you are answering to a regulator, a government customer or a security questionnaire that has started asking about post-quantum readiness, this is the strongest single argument for Gen12 and it is not really a performance argument at all. If nobody is asking, iLO 6 remains a fully supported, capable management processor and this line does not move your business case.

The AMD side of Gen12 is single-socket

The AMD-based Gen12 servers HPE has published are the DL325 Gen12 and DL345 Gen12, both single-socket, on 5th Gen AMD EPYC. The DL345 Gen12 is a 2U single-socket machine with up to 192 cores, 24 DIMM slots and up to 6 TB of DDR5 at 5200 MT/s, iLO 7, up to 36 EDSFF drives, and up to six single-wide or four double-wide GPUs. HPE's two-socket AMD rack server remains the DL385 Gen11 at the time of writing; confirm current line status with us or with HPE before you build a bill of materials around a two-socket AMD Gen12.

The single-socket 192-core machine is more interesting than it first appears, and again the reason is licensing. Where software is licensed per socket or carries a per-socket minimum, one socket carrying 192 cores and 6 TB is a fundamentally different commercial shape from two sockets carrying the same total. Where software is licensed purely per core, it is a wash on licence and a win on power and rack density. It is worth pricing both shapes rather than reflexively quoting two sockets — particularly on hyperconverged node sizing, where node count, licence count and failure-domain size are the same decision.

When ordering Gen11 is still correct

Gen11 remains an actively sold line, not a clearance item, and there are sound reasons to specify it:

• You need machines quickly. Gen11 configurations are the ones that move from stock; newer Gen12 builds more often go through the factory. The distinction and its lead-time consequences are in stock versus BTO/CTO.
• You are extending an existing Gen11 estate. Matching an installed fleet keeps firmware baselines, spares, images and operational runbooks identical. That has real value and it disappears the moment you mix generations for no reason.
• Your workload is licence-bound, not compute-bound. If per-core licensing dominates the bill, buying more cores per socket is buying more cost. A modest Gen11 core count can be the cheapest total-cost machine on the sheet.
• The validated stack you depend on has not certified Gen12 yet. Check the hypervisor, backup, storage and appliance vendors that must sign off on the platform, not just the server vendor.

When Gen12 earns the change

• NVMe density. Twelve EDSFF bays to thirty-six is the largest single jump in the table and it cannot be retrofitted to Gen11. Storage-dense and software-defined-storage designs are a straightforward Gen12 case.
• Memory bandwidth at 1 DPC. 5600 to 6400 MT/s is a real gain for bandwidth-bound workloads — but only if you populate one DIMM per channel. At 2 DPC the generation change costs you bandwidth rather than adding it.
• Security and compliance posture. Secure Enclave, FIPS 140-3 Level 3 and NIST SP 800-208 firmware signing, where somebody is auditing you against them.
• Consolidation where you own the software licences. With licensing off the table, high core counts do exactly what they promise and the consolidation ratios are excellent — the sizing method is in server refresh and consolidation.
• Direct liquid cooling, if rack power and thermal limits are the constraint on your estate rather than compute.

Frequently asked questions

Is Gen12 always faster than Gen11?

No. Per socket and on the right processor variant, yes — HPE claims up to twice the average general-compute performance versus previous generations. But a low-end Gen12 part is not faster than a high-end Gen11 part, and an E-core Gen12 configuration can perform worse than a P-core Gen11 configuration on latency-sensitive, thread-hungry workloads. Compare specific processor SKUs, not generation labels.

Does Gen12 support more memory than Gen11?

Not on the DL380. Both generations have 32 DIMM slots, 16 per processor across 8 channels at 2 DIMMs per channel, and both reach 8 TB with 256 GB modules. Gen12 runs that memory faster, but read the population: 6400 MT/s at one DIMM per channel, falling to 5200 MT/s at two. Gen11 reaches up to 5600 MT/s at 1 DPC on supported processors, so a fully populated Gen12 at 5200 MT/s is slower than a lightly populated Gen11. Capacity is not a reason to change generation here, and speed is only a reason if you populate for it.

Why would 144 cores be a problem?

Because most enterprise software is licensed per core. The 144-core figure comes from Intel's E-core Xeon 6 parts, which have no simultaneous multithreading, so 144 cores yields 144 threads and 144 licensable cores. A P-core part with fewer cores delivers two threads each and costs less to license. Under VMware, Nutanix or a per-core database licence, the dense part is frequently the more expensive machine to operate.

Can I mix Gen11 and Gen12 in the same cluster?

Generally yes at the hypervisor level, using processor compatibility modes that mask newer instruction sets so live migration works. The cost is that the cluster runs to the older baseline, so you pay for Gen12 and run at Gen11 feature level. Check your hypervisor's compatibility documentation, and treat mixed clusters as a transition state rather than a design.

Is there a two-socket AMD ProLiant in Gen12?

HPE's published AMD Gen12 rack models are the DL325 Gen12 and DL345 Gen12, both single-socket. The two-socket AMD rack server is the DL385 Gen11 as of this writing. If your design assumes a two-socket AMD Gen12, confirm current line availability before committing a bill of materials — and consider whether a single socket with 192 cores meets the requirement, since it often does and licenses differently.

Do the part numbers change between generations?

Yes, comprehensively — chassis, processor kits, memory kits, risers and drive cages are all generation-specific, and Gen11 options do not carry across. Regional suffix rules and the factory-integrated-option constraint apply the same way on both. See HPE part numbers decoded before you reuse a Gen11 configuration as the basis for a Gen12 quote.

Should we lease or consume Gen12 rather than buy it?

That is a separate decision from the generation, and it is worth separating deliberately. The mechanics of consumption pricing, what it charges for and where it beats ownership are in GreenLake versus owning the hardware. Settle the configuration first; the financing model does not change which processor variant is right.

Related

• How to choose HPE ProLiant servers and storage — platform selection and host sizing
• 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
• HPE DL360 vs DL380 Gen12 — 1U or 2U, and why memory is not the reason
• HPE part numbers decoded · sourcing HPE · GreenLake vs owning
• Lenovo ThinkSystem V3 vs V4 — the same Xeon 6 core-variant decision on another badge
• Memory population rules — why DIMM count changes bandwidth, not just capacity
• Hardware selection after the VMware licensing change · refresh and consolidation · HCI node sizing
• Dell vs HPE · Lenovo vs Dell vs HPE
• Stock vs BTO/CTO · buying 20–60 servers as one project
• 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 DW or eight SW GPUs; direct liquid cooling; iLO 7)
• HPE — ProLiant DL380 Gen11 QuickSpecs (5th Gen Xeon Scalable, up to 64 cores per socket; 32 DIMM slots, 16 per processor, 8 channels, 2 DPC; 8.0 TB at 4800 MT/s; iLO 6)
• HPE — ProLiant DL360 Gen11 QuickSpecs (32 DIMM slots; "32 x 256 GB RDIMM @ 5600 MT/s or 4800 MT/s")
• HPE — ProLiant Compute DL345 Gen12 (single socket, 5th Gen AMD EPYC up to 192 cores, 24 DIMM slots, up to 6 TB DDR5 at 5200 MT/s)
• HPE — ProLiant Compute Gen12 Embedded Security QuickSpecs (Secure Enclave with FIPS 140-3 Level 3; Silicon Root of Trust)