Xeon vs EPYC: choosing server processors for the enterprise
Intel Xeon and AMD EPYC are both x86-64, both run every mainstream operating system and hypervisor, and both are offered in servers from Dell, HPE, Lenovo and Supermicro. So the choice is almost never about compatibility. It is about how much work you can fit per socket, how much I/O and memory bandwidth you get, how much power you have to feed and cool — and, very often decisively, how much your per-core software licences will cost on top. This guide compares the two on each of those axes and sets out when each one wins.
Current top-of-stack, side by side
Comparing the latest generations per socket, from the manufacturers' datasheets and Intel ARK:
| Specification | Intel Xeon 6 6900 (Granite Rapids, P-core) | AMD EPYC 9005 (Turin) |
|---|---|---|
| Max cores / socket | 128 P-cores (Xeon 6980P); the E-core line reaches 288 (6900E) | 128 Zen 5 (9755); 192 Zen 5c (9965) |
| Max L3 cache | 504 MB (6980P) | 512 MB (9755); 1,152 MB on the previous-gen Genoa-X 9684X |
| Memory | 12-ch DDR5-6400, MRDIMM to 8800 MT/s | 12-ch DDR5-6400 |
| PCIe | PCIe 5.0, 96 lanes | PCIe 5.0, 128 lanes (up to 160 in a 2-socket server) |
| Max TDP | 500 W (6980P) | 500 W (9965 / 9755) |
| Socket | FCLGA7529 | SP5 |
Cores vs clock vs cache vs TDP
Cores. AMD has led raw core density for years and still does at the extreme — 192 Zen 5c cores per socket versus Intel's 128 performance cores. Intel answers density with its E-core ("Sierra Forest") line, which reaches 288 efficiency cores per socket. More cores means more virtual machines and containers per box — but every one of them may be a licence, which is where this decision usually turns (below).
Clock. For workloads bound by single-thread speed — some relational databases, EDA, financial and licence-limited applications — both vendors ship high-frequency bins. AMD's F-suffix Turin parts reach up to 5.0 GHz max boost (9575F, 9175F); Intel offers frequency-optimised SKUs across the range.
Cache. Both now carry very large last-level caches. AMD's 3D V-Cache (X) parts are in a class of their own for cache-bound work — the Genoa-X 9684X holds 1,152 MB of L3 — while Intel's Xeon 6 flagship reaches 504 MB.
TDP. Modern top bins hit 500 W per socket on both sides. That is a real cooling, UPS and running-cost line item (see the South Africa notes). Newer parts do more work per watt, but their absolute draw is higher, so size backup power for the peak, not the average.
Memory and I/O
The current generations have converged on 12 DDR5 channels per socket on both sides, so raw memory capacity and bandwidth are closely matched — with Intel's 6900 platform adding MRDIMM support (up to 8800 MT/s) for bandwidth-bound workloads.
I/O is where the platforms differ most. EPYC provides 128 PCIe 5.0 lanes per socket (up to 160 in a two-socket Turin server) against Intel's 96 lanes on the Xeon 6 6900 platform. For GPU-dense AI servers and NVMe-heavy storage nodes, those extra lanes translate directly into more accelerators or drives per box — a long-standing EPYC advantage that has held since the first generation.
Intel's counter is on-die acceleration: AMX (Advanced Matrix Extensions) for AI inference, introduced with 4th-Gen Sapphire Rapids and carried forward, plus QAT, DSA and DLB offload engines on the relevant SKUs.
The licensing maths — often the real decider
Most enterprise software is licensed per physical core, so the CPU you choose sets the software bill, not just the hardware bill. The three that catch buyers out:
| Software | Licensing basis | Minimum |
|---|---|---|
| VMware vSphere / Cloud Foundation (post-2024, Broadcom) | Per physical core, subscription | 16 cores per CPU |
| Windows Server 2022 (Standard / Datacenter) | Per physical core | 8 per processor and 16 per server; sold in 2-core packs |
| SQL Server 2022 | Per core | 4 per physical processor; sold in 2-core packs |
The consequence is blunt: a 192-core EPYC or a 128-core Xeon multiplies your VMware, Windows and SQL costs accordingly. For a licence-heavy estate — VMware plus SQL Server Enterprise, say — fewer, faster cores can be cheaper overall than a maximum-core part, even when the CPU itself costs more. High-frequency Intel bins and AMD F-series parts exist largely for this reason. Size the cores to the licence model, then buy the CPU.
When each one wins
| Your priority | Leans | Why |
|---|---|---|
| Maximum VM / container density | EPYC | Up to 192 cores per socket; 128 PCIe lanes |
| GPU- or NVMe-dense servers (AI, storage) | EPYC | 128 PCIe 5.0 lanes per socket (up to 160 in 2P) |
| Cache-bound HPC / EDA / technical computing | EPYC (X-series) | 3D V-Cache up to 1,152 MB of L3 |
| Licence-dominated per-core databases | Either high-frequency bin | Fewer fast cores cut licence cost — Intel frequency bins or AMD F-series |
| Built-in AI inference / offload acceleration | Intel | On-die AMX (since Sapphire Rapids), plus QAT / DSA / DLB |
| Four- or eight-socket scale-up (large in-memory) | Intel | Database & Analytics H SKUs are the parts that scale to 4S / 8S |
| Drop-in upgrade on a board you already own | Match the socket | AMD SP3 (Rome → Milan) and SP5 (Genoa → Turin); Intel FCLGA4677 (Sapphire → Emerald) |
Buying in South Africa
- New versus refurbished. The refurbished market is Intel generations 1–3 (Skylake / Cascade Lake / Ice Lake) and AMD Rome / Milan; new builds are Intel 4th–6th Gen and AMD Genoa / Turin. A refurbished dual-socket Ice Lake or Milan server delivers very high core counts at a fraction of new-part cost — an excellent fit for virtualisation where you are not licence-bound per core.
- Power and load-shedding. TDP is a running-cost and backup-sizing question, not a spec-sheet footnote. Two 500 W processors, plus DIMMs and any GPUs, is a serious UPS and generator load. Newer parts give better performance-per-watt, but plan cooling and backup around peak draw.
- Let the licence lead the sizing. For VMware and SQL estates, the software frequently costs more than the server. Cost the licences first; a mid-core, high-frequency part is often the cheaper total system.
- Plan the upgrade path within a socket family. SP3 and SP5 on AMD, and FCLGA4677 on Intel, each let you step up a CPU generation without changing the board — a cheap way to extend a fleet's life.
- Compatibility is not the constraint. Both platforms are fully supported across mainstream OEM servers, operating systems and hypervisors. The decision is workload fit, I/O, power and licensing — not whether your software will run.
Server processors and memory, priced in rand:
The Server Hub briefing
South African IT hardware news, once a week.
What’s new, what it costs in rand, and what it means for the kit you run — servers and storage, networking, backup power, surveillance and print. Every claim checked against a named source.