How much RAM for a Garry's Mod server in 2026?

A 40-player DarkRP with 120 addons often exceeds 5 Go after two map changes. For 2026, start with 2 Go for a small clean Sandbox, 4 Go for a loaded TTT or Murder, 6 Go for a serious DarkRP, and 8 Go when roleplay adds a lot of vehicles, jobs, Lua scripts and Workshop content. Beyond that, RAM almost never fixes a bad tickrate, a Lua loop, or a map that is too heavy.

Diagnosing RAM saturation on Garry's Mod

A full RAM does not always look like classic lag. On Garry's Mod, the first clear sign is often a slow memory increase after each map change, then a hard stop with Lua Panic! Not enough memory, Killed, or a restart with no clear clue in the console. If the server holds 66 tick at the start of the evening then drops to 35 tick when 30 players place props, also check the CPU. RAM does not reduce the execution time of a Lua hook run 600 times per second.

The Garry's Mod dedicated server runs through srcds_run. Workshop-mounted content, files in garrysmod/addons/ and the settings in server.cfg greatly change consumption. A blank server on gm_construct often sits between 1,1 and 1,4 Go after a few minutes. A TTT2 with 50 addons quickly reaches 3 Go. A DarkRP with 100 to 150 addons, vehicles and loaded jobs goes beyond 5 Go without trying.

UDP port 27015 plays no role in RAM. The number of slots does matter, through entities, weapons, avatars, inventories and Lua timers. A connected player rarely costs 100 Mo on its own. Ten players spamming props on rp_downtown_v4c can however add several hundred Mo, because the map, physics and scripts keep references in memory.

Choosing RAM according to the Garry's Mod game mode

The right plan depends less on the mode name than on three numbers: simultaneous players, number of active addons, map size. A Workshop collection of 8 Go on disk does not take 8 Go of RAM, but it can load 800 Mo to 2 Go if it contains a lot of models, sounds, vehicles and server-side scripts. The value host_workshop_collection in server.cfg must point to a clean collection, not to an attic where every addition from 2021 is still mounted.

Leave sv_hibernate_think at 0 if you want to save cycles when the server is empty. Setting it to 1 to keep timers active with zero players can make sense for some economy scripts, but you pay for it with ongoing usage, often 200 to 600 MB more over time with chatty addons.

Measuring RAM usage without kicking players

Measure it live. A reading taken right after a restart almost always lies, because models, Lua files, Workshop files and SQLite databases are not all loaded yet. Wait 30 minutes, one map change and at least 60 % of your slots filled. On a 40 slots server, that means 24 players connected. If your real peak is on Sunday at 21 h, measure on Sunday at 21 h.

1. Record the RAM used at startup, before any player connects. Also note the map, for example rp_downtown_tits_v2 or gm_construct.
2. Check the value after 10 players, then after 20 players. A linear increase of 30 to 60 MB per player is still normal on many RP modes.
3. Change maps once. If memory never goes back down and adds 300 to 700 MB on each cycle, look for an addon that keeps Lua references.
4. On Linux, compare RSS and VSZ. RSS is the actual resident memory. VSZ also shows reservations, so it looks scary for nothing.
5. Keep 15 to 20 % headroom. At 95 % RAM used, swap turns a 200 ms freeze into several seconds of stoppage.

ps -o pid,rss,vsz,cmd -C srcds_linux

The Starmine panel shows real-time memory, the history after restart and the peaks per instance, which avoids running this command by hand and confusing RSS, system cache and swap. The same tracking remains accessible from notre panel when you test a new Workshop collection.

Reducing RAM usage without removing half the server

The profitable cleanup starts with duplicates. Two TDM vehicle packs and two Photon packs can load the same models, the same textures and record the same entities. On a 45 players RP, removing 20 redundant addons often saves 500 MB to 1.5 GB. Removing a single large maps pack that is never used sometimes changes nothing, because a map that is not loaded mostly takes up disk space.

Lua files run server-side cost more than players think. A script placed in lua/autorun/server runs at every startup. A timer that goes through all players every 0.1 second runs 600 times per minute on 60 players. RAM goes up when this script stores tables by SteamID without clearing them on disconnect. Look for addons that create inventories, log histories, bank systems, persistent vehicles and dynamic jobs.

Do not confuse FastDL and RAM. A materials folder of 6 GB sent to players does not load 6 GB server-side. On the other hand, an addon that preloads sounds, registers 300 SWEP weapons and adds 1,200 spawnable entities increases memory from startup. Keep one production Workshop collection and one test collection. Add 5 addons, not 40. Restart, measure after 30 minutes, then validate. It is slow. It is less slow than a rollback on a wipe evening.

Add RAM or fix the CPU on Garry's Mod

The classic bad decision is to go from 6 GB to 12 GB when the server drops to 28 tick. Garry's Mod likes CPU single-core performance, because many Source and Lua calculations stay concentrated on one main thread. At 66 tick, the server has about 15.15 ms to calculate one frame. At 33 tick, it has about 30.3 ms. If your frames take 40 ms with 35 players, adding 4 GB will not bring tick back to 66.

Increase RAM when memory goes above 80 % during the real peak, or when swap appears. Stay conservative if memory stays below 65 %. The offers shown on our rates must be chosen based on your game mode and your peaks, not the compressed size of the Workshop. A well-maintained 32 slot TTT server runs better with 4 GB and good cleanup than with 8 GB and 120 forgotten addons.

Sizing a Garry's Mod server for 2026

Start from a baseline, then adjust with measurements. 2 GB cover a private Sandbox, a Build with friends, or a small Murder up to 12 players if the collection stays under 30 active addons. 4 GB cover most TTT, Prop Hunt and fun servers between 16 and 32 players. 6 GB become the proper floor for a public DarkRP with 32 to 48 slots, local SQLite database, logs, jobs, vehicles and inventories. 8 GB are for heavy RP, large maps and peaks of 50 to 64 players.

The database changes the game server RAM very little. A 300 Mo sv.db file mainly takes disk space, but addons that load everything into memory at startup can add 200 to 800 Mo. With MySQLOO or tmysql4 to MariaDB, memory usage shifts more to the database side, often 100 to 300 Mo for a small server. The bad case is when an addon loads all transactions, all vehicles and all inventories into a Lua table instead of paging them.

• Keep 15 % of RAM free at the evening peak. Below 5 %, the slightest map change becomes a gamble.
• Add 1 GB when a Workshop addition makes memory rise by 700 Mo or more after a full test.
• Do not pay for 10 or 12 GB to hide a tickrate of 33 caused by a log script running every 0,05 second.
• Test wipes, new jobs and vehicle packs on a separate instance before Friday evening.

Guides

My Garry's Mod server has 8 GB of RAM but still lags, why?

RAM mainly prevents crashes, swap and freezes linked to memory. If the tickrate drops while RAM stays below 70 %, the problem is more often the CPU, a heavy map or a Lua addon that runs too often.

How much RAM do you need for a DarkRP server with 40 players?

Plan for 6 GB for a DarkRP server with 40 players and 90 to 140 active addons. Move to 8 GB if you add a lot of vehicles, persistent inventories, custom jobs, logs and a large RP map.

Do Workshop addons use as much RAM as their displayed size?

No. A 10 GB collection can use 1 to 3 GB of RAM depending on what is actually loaded on the server side. Models, sounds and textures mostly matter during download, while Lua scripts, entities, weapons and persistent systems cost more server memory.

Should you restart a Garry's Mod server every day?

For a loaded RP server, a daily restart outside peak hours is still healthy, especially after several map changes. If RAM increases by 300 Mo on every cycle without going back down, the restart is hiding a leak, not fixing it.