Shared and Dedicated vCPU: Which One You Actually Need

Most providers sell at least two classes of compute: a cheaper shared tier and a more expensive dedicated one. The names suggest dedicated is simply better. The real difference is consistency, not speed.

Shared vCPU

On a shared instance, the physical core behind your vCPU is also serving other tenants. The hypervisor schedules between them. Most of the time there is capacity to spare and you get close to full performance.

The catch is the moment when neighbours are busy at the same time as you. Performance dips, and you have no visibility into why. The industry term for this is noisy neighbours, and it is a real effect rather than folklore.

Some shared tiers add a burst model: you accumulate credits while idle and spend them while busy. That works beautifully for spiky workloads and fails badly for sustained ones, because once the credits are gone you are throttled to a baseline that can be a small fraction of the advertised capacity.

Dedicated vCPU

A dedicated instance reserves the physical core for you. Nobody else is scheduled on it, there are no burst credits to exhaust, and performance is flat and predictable.

It costs more because the provider cannot oversubscribe the hardware.

Choosing between them

Shared is right when:

• Load is intermittent — development machines, low-traffic sites, internal tools.
• Occasional variability is tolerable.
• Budget matters more than the last few percent of consistency.

Dedicated is right when:

• CPU utilisation is sustained above roughly half, where burst models break down.
• Latency consistency is part of the requirement, such as a user-facing API with a response-time target.
• You are benchmarking and need repeatable numbers.
• A database’s performance underpins everything else.

The diagnostic

If performance is good most of the time and occasionally poor for no reason you can find in your own metrics, contention is a strong candidate. On Linux, steal time is the specific counter: it reports the proportion of time the processor was ready to run your work but the hypervisor was serving someone else. Consistently non-trivial steal time is the clearest evidence that a dedicated tier would help.

If steal time is near zero and performance is still poor, the bottleneck is elsewhere — usually memory or storage, as discussed in how to size your first cloud server.