Skip to content

Dear customers! Please send orders and inquiries to [email protected] or to Telegram @servix_uz instead of calling by phone: due to the high workload, the operator cannot answer all incoming calls.

Servix

Server fleet upgrade: when to move from Gen10 to Gen11 or Gen12

Servix technical team · Updated

In brief. It is time to move from Gen10 to Gen11 or Gen12 when there are four signs: vendor support is ending, memory slots are full, per-core licenses cost more than the hardware, or new workloads for NVMe and accelerators are beyond what PCIe 3.0 can handle. DDR4 cannot be moved to DDR5, iLO/iDRAC licenses are purchased anew, and drives and PCIe cards usually move over. The replacement pays off when consolidating five old dual-socket servers into two new ones. Gen11 is enough for typical virtualization; Gen12 is needed for Xeon 6 with up to 144 cores and 36 EDSFF bays.

A working server is rarely replaced just because a new generation has come out. Usually the reason is more specific: memory slots have run out, the vendor has announced an end-of-support date, the license bill has grown faster than the workload. Let's look at what really changes between generations, what can be moved from the old machine and what will have to be bought anew, and by what signs you can tell that the time has come.

What changed between generations

Parameter

Gen10

Gen11

Gen12

Memory

DDR4

DDR5 up to 5600 MT/s

DDR5 up to 6400 MT/s

Expansion bus

PCIe 3.0

PCIe 5.0

PCIe 5.0

Processors

1st–2nd Gen Xeon Scalable

4th–5th Gen Xeon Scalable, up to 64 cores

Xeon 6, up to 144 cores

EDSFF drives

—

up to 12

up to 36

Management

iLO 5

iLO 6

iLO 7

The main practical meaning of these rows is not in the numbers but in the fact that a compatibility boundary runs between Gen10 and the following generations: different memory, a different bus, a different management controller. Dell has the same picture: 15th generation platforms such as the PowerEdge R750 are built on DDR4 and PCIe 4.0, while the 16th generation, R760, moved to DDR5 and PCIe 5.0.

What can be moved from the old server and what cannot

Memory — no. DDR4 physically will not fit into DDR5 slots. This is the most expensive part that many expect to move over, and the most common disappointment in budget planning.

SAS and SATA drives — usually yes, but with a caveat about drive trays: the new generation has its own, and you cannot move a drive together with the old cage. Drive trays are ordered separately, and preferably in the required quantity right away.

NVMe — depends on the form factor. U.2 drives move over, but the EDSFF E3.S bays, of which Gen12 has up to thirty-six, are a different physical connector and different drives.

Expansion cards — as a rule, yes. PCIe is backward compatible, so network adapters and HBAs usually work. But check two things: whether the card is on the new server's compatibility list and whether it will become a bottleneck — a PCIe 3.0 adapter in a fifth-generation slot runs at its own speed, not the slot's speed.

Management licenses — no. iLO Advanced and iDRAC Enterprise are tied to a specific server. A license for the new machine is purchased anew, and it is usually forgotten in the estimate.

Four signs that it is time

End of service support is approaching. After this date, the vendor stops releasing microcode updates and security patches, and the service contract cannot be renewed. How end-of-sale and end-of-support dates work is covered in the article on refurbished servers.

All slots are occupied. If memory is already at its maximum and virtual machines are constrained by it, there is nothing left to add — the only option is to change the platform.

Licenses cost more than the hardware. You need to count them together: on old processors with a low core count, licensing minimums eat up the budget while less useful work gets done. A detailed calculation is in the article on choosing processors for virtualization.

New workloads have appeared that the platform cannot handle. Inference accelerators, dense NVMe pools, memory bandwidth requirements — these are the reasons the bus was changed and DDR5 was adopted. On PCIe 3.0, a modern card runs at half its capability.

When you do not need to upgrade

The workload is stable, vendor support remains in effect for another couple of years, and the servers have free memory slots and drive bays — in this situation it is cheaper to expand the existing machines. Adding memory and drives costs many times less than buying a new platform, and the result for users will be the same.

A separate case is when an upgrade is planned “just in case,” without any calculation. The specification of the new generation looks convincing, but if the current servers are loaded to a third of their capacity, the new platform will simply sit idle more efficiently.

Consolidation: where an upgrade really pays off

The most convincing scenario looks like this: five old two-socket servers are replaced by two new ones. The savings come from several items at once — fewer licenses, fewer rack units, fewer switch ports, lower electricity and cooling bills. In this case the project pays for itself in a couple of years, and this is the only case where replacing working equipment is justified economically and not just technically.

To calculate honestly, you need three figures: the actual processor and memory utilization on the current nodes, the cost of licenses for the old and new configurations, and power consumption. The first is taken from the monitoring system over a month, the second is calculated from the number of cores, and the third is taken from the rated specifications, adjusted for actual load.

A replacement sequence that keeps everything running

A typical mistake is to bring in the new servers and start the migration the same evening. The correct sequence looks different and takes more time, but it does not end with a rollback in the middle of the night.

Compatibility first. Check that your hypervisor version supports the new platform: VMware and Hyper-V have lists of certified hardware, and an older version of ESXi may simply not see the new controller or network adapter. Updating the hypervisor is a separate project that is better done before the migration, not during it.

Then firmware. A new server arrives from the factory with the firmware it had at the time of assembly, and it is almost always older than the current release. Microcode, BMC and controller updates are done before commissioning — afterwards there will be no window for them.

Then one node. Add the first new machine to the cluster and keep part of the workload on it for a week. This brings out driver and configuration incompatibilities that are not visible on a test bench.

And only then the rest. The old nodes are decommissioned one at a time, with virtual machines moved by live migration. Do not rush to dismantle the last old server: let it stay in the rack powered off for a couple of weeks as insurance.

Which to choose: Gen11 or Gen12

For typical virtualization, branch consolidation and business applications, Gen11is enough: the platform is mature and cheaper, with the same memory headroom and the same PCIe 5.0 bus. Gen12 makes sense where you need very high core density, many NVMe drives in the EDSFF format, or headroom for accelerator workloads and a long lifecycle.

To select a configuration to replace specific machines and calculate it together with licenses, write to us. Current platforms are in the Serverssection.

Related collections and tools: HPE ProLiant Gen11, Dell PowerEdge, Server for virtualization, Server configurator.

Questions and answers

Can new processors be installed in an old server?

Within one generation — sometimes yes, if the motherboard supports the required socket and a suitable BIOS version is available. Between generations — no: the socket, chipset and memory type change. A processor upgrade within a generation makes sense when it is specifically cores that are lacking, while memory and drives are fine.

What to do with old servers after the replacement?

Reasonable options: a cold standby in case of failure, a test environment, offloading non-critical services such as file storage or video surveillance, a target for backups. Selling makes sense while the platform is still supported by the vendor — after support ends, its price drops.

How long does migration take?

Technically, moving virtual machines between cluster nodes happens without stopping services. Most of the time goes into preparation: checking compatibility, updating firmware, configuring the network and storage. For a medium-sized cluster, plan for anywhere from several days to a couple of weeks, not a weekend.

Related reading

HPE introduced ProLiant Gen12 on Intel Xeon 6

ProLiant Gen12 on Intel Xeon 6: more cores and memory channels, new iLO, firmware protection. When a generation change is justified and when Gen11 from the catalog is the better deal.