UK’s trusted IT infrastructure partner since 2003
sales@servnetuk.com
0800 987 4111
Servnet
ConfiguratorGet in Touch
Server Infrastructure · How-To

How to spec a server in 2026: a practitioner's walkthrough

Servnet Editorial · Server Infrastructure Practice11 min read

Most server sizing mistakes happen at the spec stage — undersized RAM, overspecified CPU, wrong drive form factor, missed network throughput requirement. This walkthrough gives you the actual decision framework Servnet engineers use when specifying Dell PowerEdge, HPE ProLiant or Lenovo ThinkSystem for UK customers.

Server spec stack — top-down
5Workload profileVDI · DB · AI · file · web4ComputeCores · threads · clock · NUMA3MemoryDDR5 RDIMM, channels, capacity2StorageNVMe tier · SATA capacity · RAID1Network10/25/100 GbE · OOB iDRAC/iLO

Step 1 — Workload profile first, hardware second

Document the workload before opening a configurator. CPU pattern (sustained vs bursty), memory footprint (working set + headroom), storage IOPS + throughput, network throughput, peak vs average. Most teams skip this and end up over- or under-spec'd.

For VM consolidation: count VMs at peak + sum vCPU + sum vRAM + sum storage IOPS + sum network. Add 30-50% headroom for failover + growth.

For databases: capture peak transactions/sec, working set size, log throughput, backup window.

Step 2 — CPU choice (Intel Xeon 6 vs AMD EPYC 9005)

Intel Xeon 6 — per-core performance leader, best for licensing-sensitive workloads (Oracle, SQL Server per-core), AI inference (AMX accelerator).

AMD EPYC 9005 (Turin) — core count + memory bandwidth leader, best £/vCPU at typical virtualisation densities, larger L3 cache helps memory-bound workloads.

Rule of thumb: VMware vSAN / Hyper-V / typical mid-tier virtualisation → AMD EPYC. Database / SAP HANA → benchmark both.

Step 3 — Memory sizing

DDR5-6400 is current (2026). 16 / 32 / 64 / 128 GB RDIMM are the practical step sizes.

16-24 GB per VM for typical virtualisation; 64+ GB per VM for VDI gold images; 128-512 GB+ for SAP HANA / large databases.

Size to 60-70% utilisation. 90%+ memory utilisation = no headroom for vMotion / failover / growth.

Mixed DIMM sizes hurt performance — populate balanced channels.

Step 4 — Storage choice

NVMe (U.2 / U.3 / E1.S / E3.S) is the default for new servers in 2026. SAS only where SSD chassis density demands it or legacy expansion shelves are in play.

SATA SSD only for cold-tier capacity workloads where IOPS doesn't matter.

PCIe Gen5 NVMe (E3.S form factor) is the current performance ceiling — 14+ GB/s sustained per drive.

For VM workloads: typically 2-6× large NVMe drives in RAID 1/10 or HBA-passthrough to software-defined storage (vSAN, Azure Stack HCI, Nutanix AOS).

Single socket or dual?
Will any single VM need > 32 cores?
YES
Dual-socket — match licensing
NO
Single AMD Genoa = best £ / core

Step 5 — Network

10 GbE = legacy default. 25 GbE = current standard for new builds. 100 GbE = high-throughput workloads (vMotion, storage replication, AI fabric).

For hyperconverged: 25 GbE minimum, 100 GbE for production at scale.

Mellanox ConnectX-7 / 8 or Intel E810 are current dual-port 25/100 GbE workhorses.

Step 6 — Form factor + PSU + management

1U for compute-dense / minimal storage; 2U for compute + balanced storage (the workhorse); 4U for GPU / max-storage.

Always specify dual redundant PSU sized for peak draw + 20% headroom.

Always specify the management licence (iDRAC9 Enterprise for Dell, iLO 7 Advanced for HPE, XClarity Controller Enterprise for Lenovo). Without it, you can't do KVM-over-IP / firmware updates remotely.

Step 7 — Support tier

24×7×4hr on-site = standard for production. NBD = acceptable for non-production. Servnet TPM matches or beats OEM SLA at typically 30-60% lower cost — see our TPM vs OEM TCO analysis.

Key takeaways
  • Workload profile first, hardware spec second. Skipping this causes 80% of sizing mistakes.
  • AMD EPYC for typical virtualisation; Intel Xeon 6 for licensing-sensitive + AI inference.
  • NVMe is default in 2026. SAS / SATA only for specific cases.
  • 25 GbE is the new minimum standard; 100 GbE for HCI + AI fabric.
  • Always spec dual PSU + management licence + 24×7×4hr support for production.
Frequently asked

FAQs — How to spec a server in 2026

CPU + memory

How many cores do I need per VM?

Light workloads (web, app, file): 2-4 vCPU/VM. Standard business apps: 4-8 vCPU. Databases: 8-32+ vCPU. Over-provisioning vCPU is the most common sizing mistake — start lower + observe before scaling up.

How much memory per VM?

4-8 GB for light workloads; 16-32 GB for typical business apps; 64-256+ GB for databases. Size based on actual working-set memory, not "more is better".

Storage

NVMe or SAS?

NVMe for any new build in 2026. SAS only when chassis density demands it (8+ drives in 2.5" SFF) or you're expanding existing SAS-based estate.

Related

Got a question this article didn't answer?

One conversation with an engineer who's done this before. No sales script.

Talk to Servnet →