Server Maintenance Checklist: Monthly, Quarterly and Annual Tasks
Most outages were visible weeks before they happened. A degraded RAID battery, a drive logging errors since spring,...
Server memory has gone through its biggest change in a decade. DDR5 roughly doubles the bandwidth of DDR4, doubles the capacity available per module, adds error correction inside the chip itself, and is now the only option on every current Intel Xeon and AMD EPYC platform. DDR4, meanwhile, remains the memory in the vast majority of servers actually running in businesses today, and the refurbished market has made it cheaper per gigabyte than it has ever been.
That creates a question every server buyer faces in 2026: is it worth stepping up to a DDR5 platform, or does a DDR4 server still make more sense? The answer is not about preference, because the CPU decides the memory generation. It's about which generation of server to buy, and what the memory difference actually delivers for your workload.
This guide compares DDR4 and DDR5 server memory on speed, bandwidth, capacity, reliability, platform support, and cost, and gives a clear recommendation by workload.
Both are registered ECC memory in the same 288-pin DIMM form factor, but the key notch sits in a different position so they cannot be physically mixed. Beyond that, DDR5 is a redesign rather than a speed bump.
The proven workhorse. Three generations of Xeon Scalable, three of EPYC, 1.2V, 2133 to 3200 MT/s, up to 64GB per RDIMM.
The current standard. 1.1V, 4800 to 8800 MT/s, two independent sub-channels per DIMM, on-die ECC, on-module power management, up to 256GB per RDIMM.
Memory bandwidth is the figure that matters for virtualisation density, databases, and anything that streams data through the CPU. Per 64-bit module, the generations compare as follows:
Per module, DDR5-4800 is 50 percent faster than DDR4-3200 and DDR5-6400 is double. But the bigger server-side gain is channel count. A DDR4 Xeon Scalable socket has 6 or 8 channels; a current DDR5 EPYC socket has 12 and the newest has 16. Multiply the two effects and a DDR5 platform can deliver two to three times the memory bandwidth per socket of the DDR4 platform it replaces.
| Feature | DDR4 Server Memory | DDR5 Server Memory |
|---|---|---|
| Speed range | 2133 to 3200 MT/s | 4800 to 8800 MT/s |
| Bandwidth per DIMM | 17 to 25.6 GB/s | 38.4 to 70.4 GB/s |
| Channels per socket | 6 to 8 | 8, 12, or 16 |
| Channel design | One 64-bit channel per DIMM | Two independent 32-bit sub-channels per DIMM |
| Max capacity per RDIMM | 64GB (128GB LRDIMM) | 128GB to 256GB |
| Voltage | 1.2V | 1.1V |
| Error correction | Link ECC (on RDIMM) | On-die ECC plus link ECC |
| Power management | On the motherboard | On the DIMM (PMIC) |
| Cost per GB (2026) | Lowest, especially refurbished | Several times higher |
The honest summary: most business workloads are capacity-bound, not bandwidth-bound. A file server, a domain controller, or a modest Proxmox host will not notice the difference between DDR4-3200 and DDR5-4800. A busy SQL Server, an in-memory analytics engine, or a host running forty VMs will.
Both generations use ECC in server form, but DDR5 adds a second layer. DDR4 RDIMMs correct errors on the link between the memory and the controller. DDR5 adds on-die ECC, which corrects single-bit errors inside each memory chip before the data ever leaves it. The two layers are complementary: on-die ECC handles the increased error rate of denser, smaller-process chips, and link ECC still protects the path to the CPU.
| Memory Type | Error Correction | Suitability |
|---|---|---|
| DDR4 UDIMM (non-ECC) | None | Desktops only |
| DDR4 ECC RDIMM | Link ECC | Standard for all DDR4 servers |
| DDR5 ECC UDIMM | On-die ECC plus link ECC | Entry servers and workstations |
| DDR5 ECC RDIMM | On-die ECC plus link ECC | Standard for all DDR5 servers |
On-die ECC is not a replacement for proper ECC. Consumer DDR5 has on-die ECC too, but without the link ECC of a registered module it cannot report corrected errors to the system or correct errors in transit. For a server, ECC RDIMM remains the requirement in both generations.
The processor and motherboard fix the memory generation. There is no choosing DDR5 for a DDR4 server, or the reverse. In practice this means the DDR4 versus DDR5 question is really the question of which server generation to buy.
| Server Generation | Processor | Memory | Channels |
|---|---|---|---|
| Dell 13G (R630, R730) / HPE Gen9 | Xeon E5 v3 / v4 | DDR4-2133 / 2400 | 4 |
| Dell 14G (R640, R740) / HPE Gen10 | Xeon Scalable 1st / 2nd gen | DDR4-2666 / 2933 | 6 |
| Dell 15G (R650, R750) / HPE Gen10 Plus | Xeon Scalable 3rd gen (Ice Lake) | DDR4-3200 | 8 |
| AMD EPYC 7002 / 7003 (Rome, Milan) | EPYC 2nd / 3rd gen | DDR4-3200 | 8 |
| Dell 16G (R660, R760) / HPE Gen11 | Xeon Scalable 4th / 5th gen | DDR5-4800 / 5600 | 8 |
| AMD EPYC 9004 / 9005 (Genoa, Turin) | EPYC 4th / 5th gen | DDR5-4800 / 6000 | 12 |
| Dell 17G / HPE Gen12 | Xeon 6 (Granite Rapids) | DDR5-6400 / 8800 | 12 |
| AMD EPYC 9006 (Venice) | EPYC 6th gen | DDR5-8000 | 16 |
The dividing line: everything up to Dell 15G and HPE Gen10 Plus is DDR4. Dell 16G, HPE Gen11, and AMD Genoa onwards are DDR5. The refurbished market is overwhelmingly DDR4 today, which is exactly why DDR4 memory is so inexpensive and why a DDR4 server with a lot of it is such strong value.
Within either generation, the memory still has to match the platform's rated speed and type. A Xeon Scalable 2nd gen server wants DDR4-2933 RDIMMs; fitting DDR4-3200 works but runs at 2933. Our server RAM compatibility guide covers matching modules to a specific server.
The price gap is the largest it has been. Refurbished DDR4 RDIMMs sell for a fraction of the per-gigabyte price of new DDR5, and the gap widens at larger capacities. For a business that needs a lot of memory and moderate bandwidth, that gap buys a great deal of DDR4.
DDR4 is the memory of the refurbished market and DDR5 is the memory of the new-hardware market. If you've decided on refurbished, you've decided on DDR4, and the right move is to fill the server generously because the memory is cheap. If you're buying new, DDR5 is the only option and the question is simply how much.
Server memory is moving in three directions at once: faster modules, more channels per socket, and memory attached over the PCIe bus.
DDR4 RDIMM
3200 MT/s
Up to 64GB per module
Installed BaseDDR5 RDIMM
4800 to 6400
Up to 128GB per module
CurrentDDR5 MRDIMM
8800 to 12800
Up to 256GB, Xeon 6 and EPYC 9006
EmergingCXL Memory
PCIe 5.0 / 6.0
Pooled and shared across servers
EarlyChannel count keeps rising: twelve is current and sixteen has arrived on AMD's newest platform. Multiplexed RDIMMs double throughput per slot for the highest-core-count CPUs where there's no room for more DIMM slots. And CXL lets memory be attached over PCIe and shared between servers, which will eventually change how capacity is bought.
None of this changes the picture for a business buying in 2026. DDR4 will remain the memory in most running servers for years, and DDR5 is where every new platform has gone. The next refresh cycle will be DDR5; the current one can still be DDR4 if the economics favour it.
DDR5 is the better technology by every technical measure: roughly double the bandwidth per module, up to twice the channels per socket, double the capacity per DIMM, and an extra layer of error correction. For bandwidth-bound workloads, databases, analytics, and AI, a DDR5 platform is the correct choice and the only one available new.
DDR4 is the better purchase for a large share of businesses in 2026. It's the memory in every refurbished Dell 13G to 15G and HPE Gen9 to Gen10 Plus server, it costs a fraction of DDR5 per gigabyte, and most business workloads are limited by memory capacity rather than memory bandwidth. A DDR4 server filled generously with inexpensive RDIMMs outperforms a DDR5 server bought with too little memory to afford more.
Decide the server generation first, let that fix the memory type, then size capacity to the workload. Contact ServerDove for configuration assistance and memory matched to your exact server model, with delivery across Qatar. Request a Quote · Browse server memory
Most outages were visible weeks before they happened. A degraded RAID battery, a drive logging errors since spring,...
Spinning drives still run most backup targets, file servers, and NAS builds, and the market is full of...
RAID levels appear in nearly every storage decision, usually as though everyone already knows why RAID 6 beats...
A backup server isn't a production server with the workload turned down. It has inverted priorities, and Veeam...