Palworld Server Hardware Sizing: RAM, CPU and Storage by Player Count

A Palworld dedicated server at 32 concurrent players holds roughly 14–18 GB resident memory once bases are built out, and its simulation tick is single-threaded for entity movement. Those two facts drive every sizing decision below. Published system requirements mention 8 GB RAM and a 4-core CPU — numbers that hold for a four-person co-op world and for nothing else.

The sizing table in this guide comes from sustained runs on bare-metal nodes with the standard Steam dedicated server binary, measuring RSS, tick time, and save I/O at each player band. If you are choosing a host rather than a machine, the same numbers map onto the specs to look for in dedicated server offerings built for game workloads.

The Three Resources That Actually Run Out

Palworld loads every active base, every Pal AI state, and every placed structure into memory and never releases it until the world unloads. Memory therefore grows with base count and playtime, not just with connected players. A clan that has built four 50-Pal bases on a single map will consume more RAM at 16 players than a fresh world does at 40.

CPU behaves the opposite way: it is largely a function of connected players and active Pals, and it is dominated by single-thread performance. The server ticks at 30 Hz by default; when the main loop overruns 33 ms, the server announces “server FPS” drops and movement becomes visibly stuttery. Storage matters on a slower cadence — the world save writes roughly every 30 seconds by default and during shutdown — but a slow disk produces exactly the same symptom as a CPU overrun: a periodic multi-second hitch that hits every player simultaneously.

Sizing Table by Player Count

PlayersCPU (dedicated)RAMStorageTypical monthly
1–8 (friends)4 threads, 3.5 GHz+8–12 GB40 GB SSDVPS-viable
8–164c/8t, 3.6 GHz+16 GBNVMe 100 GBEntry dedicated
16–326c/12t, 3.7 GHz+24–32 GBNVMe 250 GBMid dedicated
32–648c/16t, 3.8 GHz+48–64 GBNVMe 500 GB RAID1Upper mid
64–128 (multi-world)12c/24t, 3.8 GHz+96–128 GB2× NVMe 1 TBHigh-end

The RAM column assumes bases with roughly 15 Pals each and 100+ hours of world age. Fresh worlds run 30–40% leaner. If your community rebuilds frequently or wipes seasons, subtract a tier; if you run a persistent public server with many small bases, add one.

Where Virtual Instances Start to Fail

A VPS with 16 GB of “guaranteed” RAM and four shared vCPUs can host 10–16 players comfortably. Problems appear in three places as you scale. First, vCPU steal: with 8 vCPUs on a hypervisor running 40 vCPUs of guest demand, the Palworld tick loop gets preempted and you see tick time spikes that no amount of config tuning fixes. Second, memory ballooning: overcommitted hosts reclaim guest memory under pressure, and the Palworld process gets OOM-killed mid-save.

Third, shared storage. Cloud block storage with 3,000 IOPS burst and 500 IOPS baseline will stall a 400 MB world save for two to four seconds. On bare metal with local NVMe, the same save completes in 200–400 ms. That difference is the most common reason a server “worked fine at 10 players and became unplayable at 25.”

Configuration Values That Cut Load

Most of Palworld’s default settings are tuned for small co-op groups. On a public server, raise BaseCampMaxNumInGuild limits carefully rather than leaving them at defaults that let one guild build endlessly, and lower DropItemMaxNum from 3000 to around 1500 to reduce the physics and replication load of dropped loot. PalEggDefaultHatchingTime and breeding timers do not affect server CPU and can be left alone.

  • ServerPlayerMaxNum — set to the number your hardware actually supports, not 32 by default
  • bIsUseBackupSaveData — keep true; costs disk, saves worlds
  • AutoSaveSpan — raise from 30s to 60–120s on busy servers to reduce save hitches
  • BaseCampWorkerMaxNum — cap at 15–20; 50-Pals-per-base is the largest single RAM driver
  • DeathPenalty — gameplay setting, but item-drop-on-death increases entity churn

Changes to PalWorldSettings.ini require a restart of the server process; the file is rewritten on shutdown, so editing it while running will be overwritten. Apply edits with the server stopped.

Diagnosing Tick Rate Collapse

Palworld exposes tick health through the log line reporting server FPS and through the in-game /Info admin command. Sustained values below 25 with fewer than 20 players indicate a hardware problem, not a world problem. Check top -H -p $(pgrep PalServer) for a single thread pinned at 100% — that is the tick thread and it means you need clock speed, not more cores.

If tick FPS looks fine but players report rubber-banding, the cause is usually the save write blocking the main loop on slow storage, or memory pressure forcing swap. Watch si/so in vmstat 1; any sustained swap activity on a Palworld node is a fatal configuration and the fix is more RAM. The wider diagnostic sequence is in our Palworld stability and restart scheduling notes.

Restart and Backup Discipline

Palworld leaks memory slowly across long uptimes, in the range of 300–800 MB per 24 hours on a 32-player server. Schedule a restart every 12–24 hours during low-traffic hours, and take a copy of the save directory before each restart so a corrupted save is never the sole copy. A simple systemd timer plus rsync to a second disk is sufficient; off-host copies belong in object storage.

Verify restores, not just backups. A 900 MB world that takes 40 minutes to restore is a different operational risk than one that restores in four minutes, and that difference comes down to whether the save lives on NVMe with a warm page cache versus cold S3. Test a restore quarterly.

Cost and Scaling Decisions

For a public server that reliably fills 32 slots, a dedicated 8c/16t machine with 64 GB RAM and NVMe typically runs $80–150/month depending on region and bandwidth terms. A comparable VPS with dedicated vCPUs lands near the same price but with shared storage and no BIOS-level access. The dedicated machine wins on predictability; the VPS wins on provisioning speed for the first month.

Scale by world, not by slot count. Running two 32-player worlds on two mid-range nodes isolates crashes, keeps each tick loop on its own core set, and costs less than one 128 GB machine that has to be rebooted for every update. A sensible growth path is 16 players on a 16 GB dedicated entry box, 32 on a 6c/12t with 32 GB, and 64+ split across two nodes with a shared community backend.

Leave a Reply