Xeon vs EPYC vs Ryzen for Game Servers: ST/MT Ratios and Cost per Slot

“Xeon or EPYC” is not a purchasing decision — the SKU inside that family is. Two single-socket nodes can both advertise 16 cores and 128 GB, and one will hold 20 TPS at 90 Minecraft players while the other drops to 11 TPS at 60. This comparison puts measured single-thread, multi-thread, and memory headline figures for the CPUs that actually appear in dedicated hosting catalogues next to per-slot cost, so you can pick by arithmetic instead of by brand.

Benchmark table: eight CPUs buyers actually choose from

Figures below are approximate public PassMark single-thread and CPU Mark values for the SKUs as normally seen in 2026 rental catalogues. Absolute scores move with BIOS revisions and memory population; the ratios between rows are what matter for sizing.

CPU	Cores/Threads	ST score	MT score	Typical rental class	Best fit
Xeon E5-2699 v3	18/36	~1,790	~17,500	$50–80 refurb	Many light instances
Xeon Gold 6248R	24/48	~2,500	~32,000	$120–190	Mixed fleet
Xeon Gold 6448Y	32/64	~2,900	~53,000	$260–350	Dense multi-instance
EPYC 9354P	32/64	~2,850	~62,000	$240–330	Heavy web + game mix
Ryzen 7 7700X	8/16	~4,100	~31,000	$90–140	Single bound instance
Ryzen 9 7950X	16/32	~4,600	~63,000	$170–240	Best single-thread value
EPYC 4564P	16/32	~4,300	~62,000	$200–280	Clock-bound + ECC platform
Xeon 6 6716P	32/64	~3,100	~60,000	$320–420	Density with newer platform

Reading the ratios instead of the absolute numbers

Three ratios explain most of the real-world differences between these chips.

Single-thread ratio. Ryzen 9 7950X versus Xeon E5-2699 v3 is 4,600 / 1,790 = 2.57. For a tick-bound game, that multiplies directly into headroom: if the E5 node holds 45 players at 20 TPS, the same workload on the 7950X holds roughly 110–115 before the tick budget overruns. It does not hold 5x more, because RAM, network, and disk I/O cap the result first.

Multi-thread ratio. Xeon Gold 6448Y versus Ryzen 9 7950X is 53,000 / 63,000 = 0.84 — the 32-core Xeon actually loses to the 16-core Ryzen in aggregate work. If your plan is to run twelve small instances, the Ryzen node wins on both clock and cost. If you need 40 instances of 1 vCPU each, the Xeon wins on density and chassis overhead.

Cost per single-thread point. Divide monthly price by ST score. A $120 EPYC or Xeon Gold-class node at 2,500 ST costs about $0.048 per ST point. A $200 Ryzen 7950X at 4,600 ST costs $0.043. The old refurb at $60 and 1,790 ST is $0.034 — cheapest per point, but with the worst per-slot ceiling and the highest wear risk.

Where EPYC and Xeon beat Ryzen despite lower clocks

Ryzen desktop parts win single-thread benchmarks, but three hosting realities favour server platforms:

Memory channels and capacity. EPYC 9004/9005 provides 12 DDR5 channels and up to 6 TB; Xeon 6 provides 8 channels. Ryzen AM5 gives two channels and 128 GB. A 200-player Rust node with 96 GB of world state and OS page cache is simply not a Ryzen workload.
ECC and validation. Server platforms validate registered ECC across all DIMM slots. Most consumer boards accept ECC only in limited configurations and are not validated for 24/7 load at 30 °C ambient.
PCIe lane count. EPYC delivers 128 lanes, Xeon 80–88. Ten NVMe drives plus a 25 GbE NIC plus a RAID controller exhausts Ryzen’s 24 lanes immediately.

Workload mapping: four scenarios, four answers

Scenario	Bottleneck	Right buy	Why
One modded Minecraft server, 80 slots	1 main tick thread	Ryzen 9 7950X or EPYC 4564P	ST score dominates; RAM ≤64 GB
Rust, 100+ players, weekly wipes	ST + RAM + save I/O	EPYC 9354P, 128 GB ECC	Channel count and ECC matter at scale
12 mixed small instances	Total thread capacity	Ryzen 9 7950X	Best MT per dollar in a 1U
40 light public instances	Per-instance overhead + density	Xeon Gold 6448Y	Core count plus IPMI fleet tooling

The 2020 Xeon trap, quantified

Refurbished Xeon E5 hosts are marketed as bargains because the core count is high and the price is low. Run the arithmetic: an 18-core E5-2699 v3 at 2.3 GHz base delivers a per-core throughput that, in tick terms, is roughly 55–60% of a modern 4.5 GHz core. To match the tick capacity of one 8-core Ryzen 7 7700X, you need about 14 E5 cores — which is most of the node. The remaining cores are for parallel work: chunk generation, save compression, and multiple instances. It is a legitimate shape for cheap, dense, low-player servers, and a poor one for any single instance above 40–50 players.

Verify the SKU before you trust the price

Providers occasionally list a family and ship a lower SKU. Confirm with lscpu and a 10-minute load test immediately after provisioning: read sustained clock, then compare against the ST figure you paid for. A node that cannot sustain its base clock under all-core load is throttling, and the throttle costs 10–25% of tick capacity during peak hours. If you’re comparing dedicated server hosting options across providers, score each quote by ST score per dollar and by verified sustained clock, not by core count.

Decision rule in one paragraph

If a single instance must hold a high player count, buy the highest single-thread score you can afford and accept fewer cores. If you run many instances, buy total MT capacity per dollar and confirm RAM channels and lane count before density. If the workload stores more than 64 GB in memory or needs ECC, skip consumer Ryzen and take a server platform. Apply those three rules and most of the Xeon-vs-EPYC-vs-Ryzen debate resolves itself in a spreadsheet.