ARK: Survival Ascended Server Hardware and Config Tuning for 10–100 Players

ARK: Survival Ascended (ASA) is Unreal Engine 5 with a persistent, fully-simulated dinos-and-structures world, and its server cost profile looks nothing like a session shooter. Two resources dominate: a single-thread CPU budget that grows with tamed and wild creature count, and a memory footprint that grows every day the map stays up because structures, corpses, and dropped items all persist. Plan for day 14, not day 1, or you will be restarting a map that players have spent two weeks building on.

These numbers assume vanilla-plus on a 70-slot server, with a 100-slot ceiling noted where it changes the hardware bill. For related reference hardware, see the provider’s dedicated server hosting options and match the spec table below to an actual quote.

What ASA Actually Costs in CPU

ASA runs a single authoritative game thread; the rest is noise. On a modern 3.5–4.2 GHz core (Ryzen 7000, EPYC 9004, Xeon Sapphire Rapids), measured server tick time for a vanilla-plus map lands in these ranges. Anything above ~50 ms p95 tick time shows up as rubber-banding and animal pathing stutter before players report lag.

PopulationTick p50Tick p95Recommended CPU
10–20 players, early map8–14 ms18 ms4c/8t, 3.6 GHz+
40 players, 200+ tames18–26 ms32 ms6c/12t, 3.8 GHz+
70 players, built-up map28–40 ms46 ms8c/16t, 4.0 GHz+
100 players, heavy structures40–58 ms70 ms+12c+, 4.2 GHz, or split maps

Two ASA-specific multipliers matter more than thread count. Boss fights and large breeding operations spawn hundreds of AI actors for short bursts, and each tamed dino carries a pathing and stat-recalc cost that never fully sleeps. A 70-player server with 400 tames behaves like a 100-player server with 80. If your community breeds heavily, size one tier up.

RAM: The Growth Curve Is the Requirement

The base ASA map loads at about 8–10 GB resident. That is the number most host marketing quotes, and it is the number that gets your server OOM-killed on day 10. Observed growth for a 70-slot vanilla-plus server:

DayRSS (70 slots)RSS (100 slots)Notes
1 (fresh wipe)9–11 GB12–14 GBEmpty map, starter bases
714–17 GB18–22 GBStructures, tames, drops
1418–23 GB24–30 GBWhere most OOM incidents land
30+22–28 GB30–38 GBNeeds scheduled restart

Provision 1.5× the day-30 figure as physical RAM so the kernel has page cache for .arktribe and .arkprofile writes. A 70-slot server wants 32–40 GB; a 100-slot server wants 48–64 GB. Running 32 GB on a 70-slot map works for two weeks and then fails during a raid, which is the single worst time for a restart.

Storage: Save-Write Amplification

ASA saves the full world state on a timer plus on any significant event (tame, structure placement, player logout). Those saves are multi-hundred-megabyte writes that serialise with gameplay if the disk stalls. On NVMe with 4k random write latency under 200 µs, a save adds 40–120 ms of tick jitter; on SATA SSD it is 200–500 ms; on spinning media the server is unplayable for seconds. Use NVMe and keep 20% over-provisioned. The separate-filesystem rules for heavy world saves are covered in this guide to chunk and save I/O on NVMe, which applies to any large persistent map.

The Config Values That Actually Move the Needle

  • AutoSavePeriodMinutes: default 15. Raise to 30 on a healthy NVMe box; lower to 10 only if players routinely crash during saves.
  • MaxTamedDinos: default 5000 server-wide. At 70 slots, 3000 is a reasonable cap and saves roughly 2–4 GB of RAM plus 4–6 ms tick.
  • DinoCountMultiplier: default 1.0. Dropping to 0.7 cuts wild-AI CPU by about 10–15% on a busy map.
  • bUseStructureStasisGrid: keeps distant bases frozen; essential above 50 slots.
  • Egg and mating intervals: breed-event nights spike CPU and RAM harder than boss fights; warn players and restart before them.
  • -USEALLAVAILABLECORES launch flag: lets worker threads spread, but does nothing for the main thread — do not buy cores expecting tick time to drop.

Cluster and Multi-Map Layout

ASA clusters transfer survivors, dinos, and items between maps. Each running map is a full server process with its own RAM footprint, so a five-map cluster on one machine is a memory problem before it is a CPU problem. The realistic pattern is one map per 8-core slice with 32–40 GB each, hosted on a high-memory node (192–384 GB), with cluster transfer over the local loopback or a private VLAN rather than across the public internet. Cross-map transfer latency below 2 ms eliminates the “upload timed out” failures players complain about on shared hosts.

Network and Slot Budget

ASA is not a bandwidth-heavy title. At 70 players expect 12–25 Mbps outbound sustained, with 60–100 Mbps spikes when a large tribe loads into render range of a big base. The real constraint is UDP socket buffer depth: peeroutbound traffic spikes when many players enter one area, and default net.core.rmem_max on modern kernels drops packets before the game notices. Raise socket buffers to 8 MB and verify counters daily. Bandwidth contracts of 1 Gbps unmetered are ample for a single map; 10GbE only matters for clusters and for absorbing game-update downloads.

A Reference Build for 70 Slots

ComponentSpecWhy
CPU8c/16t, 4.0 GHz+ boost, high single-threadTick time is single-thread bound
RAM40 GB DDR5 ECCDay-30 footprint plus page cache
Disk2× 1 TB NVMe, RAID1, 20% OPSave writes and update staging
Network1 Gbps unmetered, DDoS scrubbingASA is a DDoS target
OSLinux via Proton/WineGE or Windows ServerCompare tick parity on your map

Operational Habits That Keep the Map Alive

Three habits prevent most ASA incidents: a nightly restart at low population with a graceful save first, an automated backup that copies the save directory to a second disk before the restart, and a RAM alarm at 80% of physical memory so you learn about a leak before the kernel does. If other titles share the box, pin and isolate ASA’s single-thread core; a compiler on that core shows up directly as tick misses.

For teams running a long-lived cluster, the arithmetic is simple: clock speed and NVMe latency set the floor of what players feel, RAM capacity sets how many weeks you get before a forced wipe, and network buffering sets whether raids go smoothly. Buy for day 30.

Leave a Reply