Server RAM

DDR4 is fitted to most used servers and is widely available. For existing systems it's the only option: no server can be moved to DDR5 by swapping modules.

DDR5 delivers roughly 50 % more bandwidth per channel (4800 rather than 3200 MT/s), and the platforms have more channels – AMD EPYC 9004 drives twelve per socket instead of eight. That pays off for memory-hungry databases, in-memory analytics and densely packed virtualisation hosts.

Optane PMem isn't DRAM but persistent memory in DIMM form, with 128 to 512 GB per module. It sits alongside ordinary RDIMMs and makes very large memory configurations possible, but only works with two Intel generations. Our guide Optane PMem: Memory Mode and App Direct has the details.

Upgrading ECC RAM: four checks before you buy

• Module type: RDIMM, LRDIMM or 3DS LRDIMM. The JEDEC label or the hardware inventory in iDRAC or iLO tells you; new modules must be the same type.
• Channels per CPU: Xeon E5 v3/v4 has four, 1st/2nd Gen Xeon Scalable six, 3rd Gen Xeon Scalable and EPYC 7002/7003 eight. Add modules so that every channel ends up populated the same way.
• CPU memory limit: standard 1st Gen Xeon Scalable parts address 768 GB per socket, 2nd Gen parts 1 TB. Only the M and L models go further.
• Capacity per module: 32 and 64 GB RDIMMs are the cost-effective range. 128 GB LRDIMMs only make sense once you run out of slots.

How much server RAM do you need?

• Virtualisation: add up the RAM assigned to all VMs and allow 10 to 20 % headroom for the hypervisor and growth. Twenty VMs with 8 GB each come to about 192 GB – in a dual-socket 2nd Gen Xeon Scalable server, for example 12 × 16 GB, one module per channel.
• ZFS and TrueNAS: ZFS uses free memory as a read cache (ARC). 32 to 64 GB is a sound start for a small NAS, and every extra gigabyte improves the hit rate. Deduplication needs many times more.
• Databases: the working set plus indexes should fit entirely in memory. More RAM often helps more than a faster CPU here.