How to choose a server CPU
Pick a processor by load profile, not by the core count on the price list. When the task parallelises badly — a game server, an accounting system, single-threaded database queries — single-core clock speed decides. When the load is many independent requests — web, containers, transcoding, virtualisation — the total thread count decides.
- Identify the profile first: single-thread latency or aggregate throughput. That is the primary criterion.
- Xeon E3 gives four high-clock cores; the Xeon E5-2680v4 gives 14 cores and 28 threads per socket.
- EPYC brings more PCIe lanes and memory channels, Ryzen brings the highest clock at modest core counts.
- Generation beats model number: a modern core does noticeably more work per cycle.
- Verify clock speed under load rather than in the datasheet: the governor and temperature cut it down.
Clock speed or core count
A core at 3.9 GHz walks a serial chain of operations roughly 30% faster than a 3.0 GHz core of the same generation. For a query that runs in one thread, neighbouring cores contribute nothing: they sit idle.
The opposite case is 200 concurrent HTTP requests or 20 virtual machines. There the 28 threads of an E5-2680v4 drain the queue better than the 8 threads of a high-clock E3, even at a lower per-core frequency. For game projects the priority is unambiguous and covered separately in hardware for a game server.
How the families differ
| Family | Strength | Typical cores | Best suited to |
|---|---|---|---|
| Intel Xeon E3 / E series | High clock, one socket, ECC | 4 cores, 8 threads | Game servers, accounting apps, small web with heavy PHP |
| Intel Xeon E5 / Scalable | Many threads, dual socket, memory capacity | 14 cores and up | Virtualisation, containers, loaded databases |
| AMD EPYC | Memory channels and PCIe lanes, core density | 16–64 cores | NVMe storage, analytics, large VM clusters |
| AMD Ryzen | Top per-core clock, price | 6–16 cores | CI builds, game servers, compilation |
Server-grade Ryzen boards usually do not take large registered memory, so builds with 128 GB and above sit on Xeon or EPYC. On the role of error correction, see the breakdown of ECC memory in a server.
How many cores the workload needs
- A web application with the database on the same box. 4–8 cores, clock first: PHP-FPM and queries bottleneck on one thread.
- PostgreSQL or MySQL running reports. 8–16 cores plus memory: parallel plans spend several cores per query. Details in the guide to a dedicated server for a database.
- A hypervisor with 10–20 VMs. 14 cores and up, 32–64 GB of memory, otherwise the scheduler starves the guests.
- Video transcoding. Count threads: one 1080p x264 stream takes about two cores at a high preset.
How to check the CPU on a live server
A datasheet shows the maximum, not what you receive. Capture the actual numbers.
# Model, sockets, cores and threads, cache size
lscpu
# The real frequency of each core right now
grep -E 'model name|MHz' /proc/cpuinfo | head -n 8
# Frequency governor: should read performance, not powersave
cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
# Synthetic measurement: single thread and all threads
sysbench cpu --threads=1 --time=30 run > /root/cpu-1t.log 2>&1
sysbench cpu --threads=$(nproc) --time=30 run > /root/cpu-all.log 2>&1
Compare the single-thread result with the multi-thread one: if the multi-thread score does not scale with core count, memory bandwidth or the thermal budget is the limit. The full method is in the guide to benchmarking a server.
The powersave governor costs up to a third of your performance on short requests. The kernel holds a low frequency and ramps up with a delay, so fast queries finish before the clock rises. Aggressive C-states and Turbo disabled in BIOS produce the same effect.
The check takes a minute: run cpupower frequency-info and compare the current frequency under load with the base clock from lscpu. If the frequency sits below base clock at full load, look at cooling and thermal throttling instead of buying a bigger CPU. The settings are covered in server BIOS/UEFI configuration.
The processors in ZevsHost configurations
The range uses two classes. The Xeon E3-1230v5 and E3-1270v6 sit in the Start plans: four high-clock cores, 16 GB DDR4 ECC, two SSDs in RAID1, from $49 per month in Germany. That machine covers game projects and applications with single-threaded hot paths.
The Xeon E5-2680v4 powers the Pro plans: 14 cores and 28 threads, 32 GB ECC, two 1 TB NVMe drives. Dedicated Enterprise US carries two of those CPUs, 64 GB of memory and four NVMe drives in RAID10 for $149, which is the configuration for a hypervisor or a loaded database. The full list is on the dedicated server hosting page.
Key takeaways
- The load profile drives the choice: single-thread latency needs clock speed, throughput needs threads.
- E3 and Ryzen win on frequency; E5, Scalable and EPYC win on thread count and memory capacity.
- Core generation matters more than the model number: per-cycle differences reach tens of percent.
- Before concluding the CPU is too small, check the governor, Turbo and temperature — free headroom hides there.