Every dedicated server listing leads with the core count, and the core count is the least informative number on it. A 44-core Xeon from 2016 and a 24-core EPYC from 2022 are not competing for the same work: one wins on threads per euro, the other on everything that happens inside a single thread. The Xeon vs EPYC decision is really four smaller questions. How fast is each core? How much memory bandwidth reaches it? How much I/O can the platform carry? And how soon do you need the machine?
This guide answers them for the eight builds NexonHost sells, using the manufacturers' own specification pages rather than benchmark folklore.
The processors, from the spec sheets
| Processor | Launched | Cores / threads | Base / max boost | L3 cache | Memory | Bandwidth per socket | PCIe |
|---|---|---|---|---|---|---|---|
| Xeon E5-2630 v3 | Q3 2014 | 8 / 16 | 2.40 / 3.20 GHz | 20 MB | DDR4-1866, 4 channels | 59 GB/s | 3.0, 40 lanes |
| Xeon E5-2690 v4 | Q1 2016 | 14 / 28 | 2.60 / 3.50 GHz | 35 MB | DDR4-2400, 4 channels | 76.8 GB/s | 3.0, 40 lanes |
| Xeon E5-2699 v4 | Q1 2016 | 22 / 44 | 2.20 / 3.60 GHz | 55 MB | DDR4-2400, 4 channels | 76.8 GB/s | 3.0, 40 lanes |
| EPYC 7702 | 2019 | 64 / 128 | 2.00 / 3.35 GHz | 256 MB | DDR4-3200, 8 channels | 204.8 GB/s | 4.0, 128 lanes |
| EPYC 9254 | Nov 2022 | 24 / 48 | 2.90 / 4.15 GHz | 128 MB | DDR5-4800, 12 channels | 460.8 GB/s | 5.0, 128 lanes |
| EPYC 9554 | Nov 2022 | 64 / 128 | 3.10 / 3.75 GHz | 256 MB | DDR5-4800, 12 channels | 460.8 GB/s | 5.0, 128 lanes |
| EPYC 9754 | Jun 2023 | 128 / 256 | 2.25 / 3.10 GHz | 256 MB | DDR5-4800, 12 channels | 460.8 GB/s | 5.0, 128 lanes |
The figures are per processor. Most of the builds below use two.
The eight builds
| Build | Cores / threads | RAM | Storage | Port options | Delivery | Monthly |
|---|---|---|---|---|---|---|
| 2× Xeon E5-2630 v3 | 16 / 32 | 128 GB DDR4 | 1× 240 GB SSD | 1–20 Gbps | Minutes | €129 |
| 2× Xeon E5-2690 v4 | 28 / 56 | 128 GB DDR4 | 1× 240 GB SSD | 1–20 Gbps | Minutes | €249 |
| 2× Xeon E5-2699 v4 | 44 / 88 | 256 GB DDR4 | 1× 240 GB SSD | 1–20 Gbps | Minutes | €425 |
| 1× EPYC 9254 | 24 / 48 | 128 GB DDR5 | 2× 1 TB NVMe, RAID 1 | 1–50 Gbps | About 2 weeks | €449 |
| 2× EPYC 9254 | 48 / 96 | 256 GB DDR5 | 2× 1 TB NVMe, RAID 1 | 1–50 Gbps | About 2 weeks | €679 |
| 2× EPYC 9554 | 128 / 256 | 512 GB DDR5 | 2× 1 TB NVMe, RAID 1 | 1–100 Gbps | About 2 weeks | €959 |
| 2× EPYC 7702, Bucharest only | 128 / 256 | 512 GB DDR4 | 2× 240 GB M.2 + 2× 960 GB NVMe | 1–100 Gbps | About 2 weeks | €989 |
| 2× EPYC 9754 | 256 / 512 | 1 TB DDR5 | 2× 1 TB NVMe, RAID 1 | 1–100 Gbps | About 2 weeks | €1,519 |
The Xeon and EPYC 9004 builds include 40 Gbps / 4 Mpps of DDoS mitigation and one IPv4 address, and none of the eight carries a setup fee.
Question 1: how fast is each core?
Clock speed is the only per-core figure both manufacturers publish, and it is a floor rather than a verdict. A core designed in 2022 does more work per clock than one designed in 2014, and neither Intel nor AMD publishes a single number for that difference, so treat any chart that claims one with suspicion. What the spec sheets do tell you is the ceiling: the EPYC 9254 boosts to 4.15 GHz, while the E5-2630 v3 tops out at 3.20 GHz.
That ceiling matters for game servers, many of which run each world's simulation on one main thread; for databases answering thousands of small queries; and for any application with a single hot lock. For that kind of work the single EPYC 9254 at €449 is the right build, even though the €425 Xeon E5-2699 v4 has nearly twice as many cores.
Question 2: how much memory bandwidth reaches each core?
Divide each manufacturer's memory bandwidth by the core count and the range sorts itself:
| Processor | Bandwidth per socket | Cores | Per core |
|---|---|---|---|
| EPYC 9254 | 460.8 GB/s | 24 | 19.2 GB/s |
| Xeon E5-2630 v3 | 59 GB/s | 8 | 7.4 GB/s |
| EPYC 9554 | 460.8 GB/s | 64 | 7.2 GB/s |
| Xeon E5-2690 v4 | 76.8 GB/s | 14 | 5.5 GB/s |
| EPYC 9754 | 460.8 GB/s | 128 | 3.6 GB/s |
| Xeon E5-2699 v4 | 76.8 GB/s | 22 | 3.5 GB/s |
| EPYC 7702 | 204.8 GB/s | 64 | 3.2 GB/s |
The EPYC 9254 feeds each core more than twice as much memory bandwidth as anything else in the range. That makes it the build for analytics, in-memory caches, large PostgreSQL or MySQL instances, and anything that scans more data than fits in cache. The EPYC 9754 sits at the other end: 128 cores per socket sharing the same twelve channels, which suits many independent tasks, such as containers, CI jobs and microservices, far better than one large one.
Question 3: how much I/O can the platform carry?
The Xeon E5 v3 and v4 processors provide 40 lanes of PCIe 3.0 each. The EPYC 9004 series provides 128 lanes of PCIe 5.0 per processor, and the 7702 provides 128 lanes of PCIe 4.0. You see the difference directly in what each build can be configured with:
- Network. The Xeon builds take ports up to 20 Gbps. The EPYC 9254 builds ship with two 25 Gbps SFP28 ports and go up to 50 Gbps. The EPYC 9554 and 9754 builds ship with two 100 Gbps QSFP28 ports and go up to 100 Gbps.
- Storage. The EPYC 9254 builds take ten extra NVMe drives; the 9554 and 9754 builds take twenty-four. The Xeon builds take up to six SSDs.
If the server is going to move a lot of data, over the network or to disk, I/O decides the build before the processor does.
Question 4: how soon do you need it?
The Xeon builds provision automatically, typically within minutes of payment. The EPYC builds carry an estimated delivery of about two weeks. If you are replacing a failing server this week, that settles it: a Xeon can carry the load now, and an EPYC can replace it when it arrives.
Which build for which workload
| Workload | Build | Why |
|---|---|---|
| Web and app servers, staging, CI runners, VPN, mail | 2× E5-2630 v3 or 2× E5-2690 v4 | Cheapest whole server; online in minutes |
| Parallel batch work on a budget: crawlers, compile farms, encoding | 2× E5-2699 v4 | 88 threads for €425 |
| Game servers, single-threaded apps, busy databases | 1× or 2× EPYC 9254 | Highest boost clock and bandwidth per core |
| Virtualisation hosts, large in-memory datasets | 2× EPYC 9554 | 128 fast cores, 512 GB DDR5, up to 100 Gbps |
| Container density, many small services | 2× EPYC 9754 | 256 cores at €5.93 each, the lowest per core |
| 128 cores in Bucharest at the lowest price | 2× EPYC 7702 | €989 with DDR4 |
The Xeon vs EPYC verdict
On single-thread speed, memory bandwidth and I/O, EPYC wins by a wide margin, as you would expect when the EPYC 9004 chips are six to nine years newer than the Xeons. On monthly price and speed of delivery, the Xeon builds win. Neither is the better processor in the abstract. The right build is the one whose weakest number is a number your workload does not care about.
If you are not sure, measure. Run the real application on a Xeon build for a few days, watch whether it runs out of cores, clock or memory bandwidth first, and choose the EPYC that fixes that particular limit. What €129 a month actually buys prices the entry build and its add-ons in detail.
Where to start
All eight builds, with current prices and locations, are on the dedicated servers page, and the bare metal overview explains what single-tenant hardware includes. If the port matters more than the processor, high-bandwidth dedicated servers cover the faster ports, which run from 1 Gbps up to 100 Gbps depending on the build.
Specifications and pricing are as of 28 September 2026 — check the product pages for current figures.
Sources
- Intel specification pages for the Xeon E5-2630 v3, Xeon E5-2690 v4 and Xeon E5-2699 v4: launch dates, cores, clocks, cache, memory, bandwidth and PCIe lanes. Retrieved 28 September 2026.
- AMD specification pages for the EPYC 9254, EPYC 9554 and EPYC 9754, and AMD's server processor specifications table for the EPYC 7702. Retrieved 28 September 2026.
- NexonHost store configurator: builds, network cards, port options, storage options and prices. Retrieved 28 September 2026.
- Computed: bandwidth per core is the manufacturer's per-socket memory bandwidth ÷ cores per processor; price per core is the monthly price ÷ physical cores.
