← All guides

Decide

Choosing and sizing a dedicated server

Technical buyers, engineers · 9 min read

The most common mistake in buying a dedicated server is sizing for the peak you imagine instead of the workload you have. Start from what the application actually does, find the bottleneck, and the rest of the spec follows. Here is a methodology that avoids both under-provisioning and paying for capacity you will never use.

Step 1: find the bottleneck

Almost every workload leans one way. Identify which, and size for that first:

  • CPU-bound: rendering, transcoding, compilation, many concurrent requests.
  • Memory-bound: databases, caches, in-memory analytics, large JVM heaps.
  • I/O-bound: write-heavy databases, log processing, large datasets on disk.

Step 2: match a profile

As a starting point, common workloads map to broadly predictable shapes. Treat this as a first approximation to refine, not a rule:

WorkloadLeans onSensible starting point
Web / API tierCPU + networkMid core count, moderate RAM, fast network
Transactional databaseMemory + NVMeHigh RAM, NVMe storage, ECC essential
Virtualisation hostCores + RAMHigh core count and RAM, room for many VMs
Analytics / batchCPU + I/OHigh cores, fast storage, generous bandwidth

Step 3: size each component

CPU: cores versus clock

More cores help concurrency; higher clock speed helps single-threaded work. Do not assume more cores is always better — a database with a few hot connections may benefit more from clock speed than from a high core count. Match the CPU to how the application actually parallelises.

Memory: size for the working set

Size RAM for your working set plus headroom, not for the total dataset. For databases and caches, keeping the hot data in memory is usually the single biggest performance lever. Insist on ECC for anything that matters — it catches memory errors that would otherwise corrupt data silently.

Storage: NVMe where latency counts

Use NVMe for latency-sensitive databases and anything in the hot path; use higher-capacity tiers for bulk and cold data. The gap between NVMe and older storage shows up directly in database transaction latency, so do not economise on the disk that holds your hot data.

Technical detail

Every Metal on Cloud server includes IPMI/KVM for out-of-band management, including BIOS-level access when the OS is down, so you can recover a machine without a site visit. Standard configurations are Dell PowerEdge R750 chassis with Intel Xeon Gold CPUs, ECC RDIMM memory, and U.2 NVMe storage — the same hardware class enterprise data centers run globally. Need a specific layout? We will build to spec rather than force you into a fixed tier.

Step 4: leave headroom, without overspending

Aim to run comfortably at your normal load with room to absorb spikes — not to sit idle 80% of the time. A useful rule of thumb is to target peak utilisation around 70–80% of capacity, which leaves margin without waste. Because provisioning is fast and billing is month-to-month, it is cheaper to start right-sized and add capacity than to overbuy on day one.

Key takeaway

Find the bottleneck, match a profile, size each component for it, insist on ECC and NVMe where they count, and leave sensible headroom. Tell us the workload and we will recommend a configuration — not the largest available specification.

Ready to talk specifics?

Get a Quote