AMD EPYC or Intel Xeon: how to choose a server platform
Servix technical team · Updated
In brief. The choice between AMD EPYC and Intel Xeon today is decided by three things, and none of them is about “who is faster”: the licensing scheme (per-socket favors EPYC with its density, per-core favors high-frequency Xeon), the nature of the workload (single-threaded accounting systems versus dense virtualization) and cluster homogeneity — live migration between Intel and AMD nodes is impossible. In the Servix catalog, the single-socket EPYC 7313 with 16 cores costs 14,376,470 UZS, and the comparable in core count Xeon Gold 6226R costs 14,149,473 UZS.
The question “Intel or AMD” in the server segment stopped being obvious several years ago. Today both platforms are present in all lines — HPE ProLiant, Dell PowerEdge, Lenovo — and the choice is determined by the nature of the task, not by habit. We look at the criteria for making the decision.
Where AMD EPYC wins
- Core density. Substantially more cores fit per socket: up to 96 in the Genoa generation, up to 192 in Turin. Where Intel needs two servers, one is often enough with EPYC.
- Single-processor configurations. An underestimated scenario: a single-socket EPYC is comparable to a previous-generation dual-processor system in core count, memory channels and PCIe lanes. The result is a cheaper server, lower power consumption and lower licensing costs where licenses are counted per socket.
- Memory channels. Twelve channels per socket versus eight for 4th–5th generation Xeon Scalable. On tasks sensitive to memory bandwidth — analytics, inference, large DBMS — this is noticeable.
- PCIe lanes. There are more of them, which matters for configurations with many NVMe drives and network adapters.
- Cache size. In some lines, L3 cache reaches 384–512 MB; on tasks where the working set fits in cache, the difference is many times over.
- Cost per core. Calculated per core, the platform is traditionally more cost-effective.
Where Intel Xeon wins
- Ecosystem maturity. The widest compatibility list with enterprise software, well-tuned drivers, predictable behavior during hypervisor updates.
- Built-in accelerators. The QAT, DSA and IAA hardware blocks offload the cores in encryption, compression and data processing — provided the software can use them.
- High-frequency models. For applications limited by single-thread speed — and that is a significant share of accounting systems, including 1C — models such as the Gold 6248R at 3.0–4.0 GHz (24,440,000 UZS) or the Gold 5317 at 3.0–3.6 GHz (14,376,470 UZS) remain a strong argument.
- Flexibility of the Xeon 6 line. The split into performance P-cores and efficient E-cores makes it possible to match the processor more precisely to the workload profile.
- Compatibility with the existing fleet. A homogeneous cluster is easier to operate: mixing platforms in one virtualization cluster limits live migration of machines.
Generations and what is in the catalog
Platform | Generation | Cores per socket | Memory | Socket |
|---|---|---|---|---|
Intel Xeon Scalable 2 | Cascade Lake | up to 28 | 6 DDR4-2933 channels | FCLGA3647 |
Intel Xeon Scalable 3 | Ice Lake | up to 40 | 8 DDR4-3200 channels | FCLGA4189 |
Intel Xeon Scalable 4–5 | Sapphire / Emerald Rapids | up to 64 | 8 DDR5 channels | FCLGA4677 |
Intel Xeon 6 | Granite Rapids | up to 128 P-cores | 8 or 12 DDR5 channels | LGA4710 / LGA7529 |
AMD EPYC 7003 | Milan | up to 64 | 8 DDR4-3200 channels | SP3 |
AMD EPYC 9004 | Genoa | up to 96 | 12 DDR5 channels | SP5 |
AMD EPYC 9005 | Turin | up to 192 | 12 DDR5 channels | SP5 |
From AMD, the catalog has the EPYC 7313 — 16 cores, 3.0–3.7 GHz, 128 MB of L3 cache on the SP3 socket (14,376,470 UZS). The Intel line is represented more widely, from the Xeon E-2224 for 2,156,470 UZS to the 32-core Platinum 8358 for 50,317,647 UZS. The full list with prices and socket mapping is in the selection server processors.
The key factor people forget: licenses
The cost of hypervisor and DBMS licenses often exceeds the cost of the server itself. The licensing scheme radically changes the calculation:
- Per-socket licensing — processors with the maximum number of cores pay off. Here EPYC gives direct savings: one license for twice the computing power.
- Per-core licensing — the situation is the opposite. “Extra” cores become expensive, and fewer high-frequency cores are more cost-effective. VMware also applies a minimum of 16 cores per processor: an eight-core chip is billed as a sixteen-core one.
- Windows Server is sold in 16-core packs with a minimum of 16 per server and 8 per processor — that is, a dense 64-core EPYC requires four packs.
Calculate the total cost of the solution, not the hardware price: over a three-year horizon, the difference between approaches can be several-fold. A detailed calculation is in the article Proxmox or VMware in 2026.
NUMA and why “more cores” is not always faster
On both platforms, the processor is internally divided into NUMA nodes: cores work faster with “their own” part of memory. In EPYC, the chiplet layout makes this more noticeable — with poor tuning, a virtual machine can get memory from another node and lose performance for no apparent reason.
Practical consequences: virtual machines should not be made wider than one NUMA node unless necessary; memory must be distributed symmetrically across all channels (see the guide server RAM); in the BIOS, check the NPS (Nodes Per Socket) mode on AMD and Sub-NUMA Clustering on Intel — they define how the processor is presented to the hypervisor.
Practical recommendations by task
Task | Sensible choice |
|---|---|
Dense virtualization, per-socket licensing | AMD EPYC, configuration with maximum cores |
1C, accounting systems, single-threaded applications | High-frequency Intel Xeon |
DBMS with per-core licensing | Fewer cores, higher frequency — both platforms, calculate by license cost |
Container platforms, microservices | Both; Xeon 6 with E-cores or high-density EPYC |
Analytics and inference, memory-bound | AMD EPYC with twelve channels or Xeon 6 6900P series |
Expanding an existing cluster | The same platform that is already installed — for homogeneity |
Compact server for a single service | Single-socket EPYC — often the best value for the capabilities |
On compatibility and migration
Claims that “AMD is less well supported” are no longer relevant today: all leading hypervisors and enterprise operating systems work with EPYC as standard. There is only one real restriction — within one virtualization cluster, processors must be from the same vendor. Live migration of virtual machines between Intel and AMD nodes is impossible because the instruction sets differ.
Hence the migration practice: a new cluster is built entirely on the new platform, and the old one is phased out gradually, moving machines via shutdown or replication. This needs to be planned in advance — together with a maintenance window and a check of hypervisor version compatibility. The replacement procedure is covered in the article on upgrading the server fleet.
What to check before purchasing
- How your hypervisor and DBMS are licensed — per socket or per core, and what minimums apply.
- The workload profile: whether it is limited by a single thread, by total cores or by memory.
- Which platform is already in the cluster — they cannot be mixed.
- Whether the selected processor is supported by your hypervisor version and the server platform (the compatibility list from HPE and Dell).
- Whether power and cooling are sufficient: higher-end EPYC and Xeon processors reach 400–500 W per socket, and the rack feels it.
We will put together configurations on both platforms for your task and compare them by total cost including licenses — write to us. Platforms are in the section Servers, processors with prices are here.
Illustration: PantheraLeo1359531, Wikimedia Commons (CC BY 4.0).
Related selections and tools: Server for virtualization, Server configurator.
Questions and answers
Can Intel and AMD servers be mixed in one cluster?
Physically — yes, they will work. But live migration of virtual machines between nodes from different vendors is impossible: the instruction sets differ, and the hypervisor does not guarantee correct state transfer. Therefore, a cluster is built from nodes of one platform, and during a transition a new cluster is built alongside.
Which is better for 1C: EPYC or Xeon?
For 1C, single-core frequency matters more than the number of cores: it has many single-threaded operations. A Xeon with a frequency of 3 GHz and above — for example, Gold 6248R or 5317 — usually gives better responsiveness than an EPYC with twice as many slow cores. If virtualization and other services also run on the same server, the calculation changes.
Is it true that EPYC is cheaper?
Per core — as a rule, yes. For the solution as a whole, you need to count licenses: with per-core licensing, a dense EPYC can cost more than a Xeon with fewer fast cores. Compare the configuration together with the software over three years.
Are Intel's built-in accelerators needed?
Only if your software can use them. QAT accelerates encryption and compression in solutions that support it, DSA accelerates memory operations, IAA accelerates data compression in databases. If your stack does not support them, the accelerators give nothing, and you should not choose a processor because of them.
Related reading
A new generation of servers: Intel Xeon 6 and AMD EPYC 9005
P-cores and E-cores in Xeon 6, EPYC 9005 core density, HPE ProLiant Gen12 and Dell PowerEdge 17G. Whether to switch now and what to consider for power and cooling.
Intel Xeon 6 or AMD EPYC 9005: which to choose for virtualization
How Intel Xeon 6 differs from AMD EPYC 9005, how core density affects the cost of VMware licenses, and which processor to choose for a virtualization cluster.
HPE ProLiant Gen11 on AMD EPYC Genoa: an overview of the DL325/DL345/DL365/DL385 line
HPE ProLiant Gen11 line on AMD EPYC 9004: DL325, DL345, DL365, DL385 and the Arm model RL300. Memory, drives, cooling, prices in Tashkent and how to choose.
HPE ProLiant DL325 and DL345 Gen12: single-socket servers on AMD EPYC 9005
HPE ProLiant DL325 and DL345 Gen12 on AMD EPYC 9005: 12 memory channels, one socket instead of two, license savings. How they differ and what to choose now.




