Minecraft Server Memory and Slot Sizing by Play Style

“How much RAM does a Minecraft server need?” is the most asked and least useful question in game hosting, because the answer depends entirely on how people play. Two servers with the same 40 online players can need 6 GB and 32 GB respectively — vanilla survival versus a 300-mod kitchen sink with chunk loaders. Memory sizing has to start from play style, then be validated against slot count, view distance, and entity density.

This breakdown covers the memory and slot math for vanilla, modded, minigame, and creative/redstone-server styles, plus the slot oversubscription traps that cause most “why is my server lagging” posts. If you are choosing hardware, compare against the dedicated game server options we track before ordering.

The Baseline: Memory Per Player by Play Style

Empirical numbers from well-run servers: vanilla survival runs 60–100 MB per concurrent player once chunks are loaded, plus 1.5 GB baseline. A 40-slot vanilla server therefore needs about 6 GB heap. Modded packs run 250–450 MB per player plus a 3–5 GB baseline for the mod registry and generated structures. A 40-slot All the Mods 9 server needs 18–24 GB. Paper/Spigot with 30 plugins sits between at 120–180 MB per player.

These numbers assume view distance 8–10 chunks. Doubling view distance roughly quadruples loaded chunks per player, and memory follows. At view distance 16, per-player memory on vanilla climbs to 250–350 MB. Most “I need 32 GB for 20 players” reports are really view-distance=16 reports — cut the view distance before you buy RAM.

Slots Versus Actual Concurrency

Slots are a licence to connect, not a load forecast. A 100-slot survival server typically peaks at 25–45 concurrent players except on launch weekend. Sizing RAM for 100 concurrent when your realistic peak is 40 wastes 6–10 GB. Conversely, a 50-slot minigame server with match queues can genuinely hold 50 in short bursts, and those bursts are what your hardware must survive.

The practical rule: size memory for 1.4× your 95th-percentile concurrent count, not your slot number. Track concurrency for two weeks before you buy. If you must size blind, use slot count × 0.5 as a conservative concurrency estimate for survival servers, and slot count × 0.85 for minigame networks where queueing is normal.

Vanilla and Paper: The 4–8 GB Band

A Paper server at 40 concurrent players, view distance 8, simulation distance 6, is comfortable in 6 GB with 4 GB free for the OS and page cache. Push to 80 concurrent and you want 10–12 GB. The gains from going beyond 12 GB on vanilla are near zero — the extra memory sits idle while GC scans a larger heap and pauses longer. Dedicated 8 GB VPS plans handle 30–50 players; 16 GB dedicated boxes handle 100–150.

Note the difference between heap and container memory. On a rented dedicated server you control both, so set -Xmx to about 75% of allocated RAM and let the kernel cache region files. On shared/VPS plans the host’s cgroup limit decides — leaving too little headroom causes the OOM killer to take the JVM out during a world save, which is the most common cause of unexplained crashes on budget hosts.

Modded Packs: 16–32 GB and the GC Trap

A 250-mod pack at 30 concurrent players holds 14–20 GB live. Allocate 24 GB heap on a 32 GB machine and you have room for the OS plus a small MariaDB for player data. At 60 players, expect 28–36 GB live and move to a 64 GB machine. Do not run modded packs at 40 slots on a 16 GB box “because it fits” — it fits until a player loads a chunk with a mob farm in it.

The GC trap: with G1GC, a 32 GB heap under modded load produces 80–150 ms pauses during young collections and 300 ms+ during mixed collections. Players see this as brief freezes when crossing chunks. Two fixes work: keep heap at 1.4–1.6× live set and force a periodic full GC during low-pop hours, or switch to ZGC if you are on Java 21+ and can accept ~10% throughput trade for sub-10 ms pauses.

Minigame Networks: Many Small Heaps Beat One Big Heap

A network running BedWars, SkyWars, and a lobby does not want one large world. It wants 12–20 small instances behind Velocity, each with 2–4 GB heap and 2 vCPU. Total memory for 200 concurrent players across 20 instances: about 60–80 GB, spread over a 128 GB machine. Per-instance GC pauses stay under 40 ms because each heap is small, and a crash takes down one arena instead of the whole network.

Slot allocation on this topology is per-instance, not per-machine. Give each arena server 24–32 slots and use the proxy to route. This is far more memory-efficient than a 200-slot monolith, and it lets you bin-pack instances onto cores — 20 instances over 16 physical cores with 4 threads each is comfortable on a modern EPYC.

Creative and Redstone Servers: The Outlier

Plot-based creative servers (PlotSquared, CreativeServer) look cheap per player but carry heavy world metadata. A 50,000-plot world uses 8–12 GB just for plot data and schematics in memory with default settings. Redstone-heavy technical servers are worse: a single lag machine can hold 40,000+ active block entities and consume 4–6 GB alone. For these, plan 32–64 GB regardless of player count.

Technical communities also need entity-count caps, which is a configuration task more than a hardware one. Set max-entity-collisions, limit redstone tick rate per chunk, and pre-generate the world so chunk generation does not compete with tick time. Pre-generation alone typically cuts peak MSPT by 20–30% on explored worlds.

Matching Slots to CPU, Not Just RAM

Memory sizing without CPU sizing produces a server that has plenty of RAM and still lags. As a rule: one modern core at 3.8 GHz+ per 40–50 vanilla players at view distance 8; one core per 20–25 modded players; one core per 30 minigame arena instances. A 40-slot modded server on 2 vCPU will MSPT-spike at 25 players no matter how much RAM you give it.

Validate with Spark: /spark profiler start for 60 seconds at peak, then read the tick breakdown. If the main thread is spending 60%+ in entity ticking, you need CPU. If it is spending time in GC, you need heap tuning. If it is spending time in chunk I/O, you need NVMe with higher IOPS — 200K random reads is the floor for a busy world. Our sizing guide on dedicated server hardware for game communities covers the storage tier matching in more depth.

Leave a Reply