Hosting and hardware
CPU selection
Section titled “CPU selection”A CPU with strong per-core performance will generally be a better choice than one with more, slower cores. Gameplay, JavaScript, spatial streaming and incoming packet processing share the main thread. Faster individual cores help that work finish within the available tick time.
Replication workers distribute outgoing connection work across additional cores. A balance of core count and per-core performance is useful for this workload. For a VM, the CPU model, dedicated or shared CPU time and sustained usage quotas can all affect the available performance.
Load test results
Section titled “Load test results”On 12 November 2026, we tested a Release server on a Windows PC with an Intel Core i5-14600KF and 32 GB RAM. The CPU has six performance cores, eight efficient cores and 20 logical processors. Intel specifications.
The server was assigned six logical processors. Simulated clients ran on the other 14, on the same machine. Server and generator resource usage were measured separately.
Workload
Section titled “Workload”| Part | Test conditions |
|---|---|
| Players and streaming | 300 connected players in groups of 128, 128 and 44, separated by 1,000 metres. Each player in a full group observed 127 others. Each player in the last group observed 43. |
| Movement | Every player moved continuously, with a 25 ms sync interval, targeting 40 updates per second. |
| Initial synced metadata | Each player had a profile containing an ID, name and level, plus a team, job and equipment list. |
| Changing synced metadata | Once per second, each player’s activity metadata was updated with a sequence number and position. Those changes were replicated to observers. |
| Script events | Each player sent one event per second with its index, sequence number and position. The server processed it and replied with an acknowledgement, totalling about 300 events and 300 replies per second. |
| Replication | Three replication workers and send_rate = 1048576 bytes per second per connection. |
This used a small test script and the test world. GTA clients, voice traffic, resource downloads and a production mod were excluded from the test.
Each run lasted 110 seconds. Measurements cover the interval from 60 to 100 seconds.
Measured resource usage
Section titled “Measured resource usage”| Server metric | First run | Repeat |
|---|---|---|
| Average CPU usage | 3.32 CPU equivalents | 3.33 CPU equivalents |
| Peak private memory | 105.3 MiB | 104.9 MiB |
| Peak working set | 129.9 MiB | 129.3 MiB |
| Game traffic received by all clients | 42.33 MB/s, about 339 Mbit/s | 42.38 MB/s, about 339 Mbit/s |
| Mean update delivery | 39.98 Hz | 39.96 Hz |
| Lowest observer average | 39.91 Hz | 39.88 Hz |
| p99 update age | 50 ms | 50 ms |
One CPU equivalent means one CPU-second of work per second of elapsed time.
For example, 3.32 represents total server CPU work across its threads, equivalent
to keeping about 3.3 logical processors fully busy. Per-thread measurements help
show whether one busy thread limits performance despite spare CPU capacity.
Private memory measures memory committed exclusively to the process. The working set measures its resident memory in RAM. The small test script and world account for the low memory figures. Your mod’s state, scripting heap and dependencies add to them.
The simulated clients consumed another 5.5–5.7 CPU equivalents. A dedicated game host only carries the server side of that workload.
The lowest observer value is that receiver’s average across its peers. Update ages are upper bounds from 10 ms histogram buckets, measured on loopback. Both runs maintained the input workload and had no failed script acknowledgements during the measurement window.
Extrapolating to more players
Section titled “Extrapolating to more players”For 500 players under the same conditions, two simple CPU projections give:
- Scaling by connected players:
3.32 × 500 / 300, or about 5.5 CPU equivalents. - Scaling by sender-to-receiver pairs in the same group layout: about 6.0 CPU equivalents.
The corresponding projections for 1,000 players are about 11–12 CPU equivalents. These estimates assume similar work per player or observer pair. Main-thread limits, worker distribution and transport costs can change the result. Additional cores help only the work that can use them.
Traffic at different sync intervals
Section titled “Traffic at different sync intervals”These projections use the same dense groups, including a final partial group,
with every player moving. MB/s uses decimal megabytes. One MB/s equals 8 Mbit/s.
| Connected players | 25 ms interval, 40 Hz | 40 ms interval, 25 Hz |
|---|---|---|
| 100 | About 0.10 Gbit/s | About 0.07 Gbit/s |
| 300 | 0.34 Gbit/s measured | About 0.23–0.24 Gbit/s |
| 500 | About 0.61 Gbit/s | About 0.41–0.43 Gbit/s |
| 1,000 | About 1.23 Gbit/s | About 0.82–0.86 Gbit/s |
Only the 300-player result at 25 ms was measured for the populations in this table. The other entries scale traffic by sender-to-receiver pairs. The 40 ms model reduces the sync portion while keeping other application traffic at its previous rate. Its range accounts for different assumptions about transport overhead. Compression ratios may also change at another interval.
The measured traffic is the sum of simulated clients’ transport receive rates. We recommend allowing headroom for real-network overhead and retransmissions, voice and other services sharing the uplink. At 1,000 densely grouped players, the 40 ms projection leaves limited room on a 1 Gbit/s link. The 25 ms projection exceeds that link’s capacity.
What changes your requirements
Section titled “What changes your requirements”Actual resource usage depends on a set of factors, including your mod’s overhead:
| Factor | Effect on the workload |
|---|---|
| Nearby players and entities | More observers create more outgoing updates. The test’s 128-player groups describe its layout. stream_distance sets a radius, so a crowded area can contain more entities. |
| Activity | Idle players usually generate less traffic. Movement, combat, vehicles and frequent streaming changes add work. |
| Metadata and script events | Payload size, update frequency and the number of recipients affect CPU work and traffic. |
| Mod logic and state | Callbacks, timers, entity processing, stored state and database integration add CPU and memory costs. |
| Voice and hosted services | Voice, databases, caches and resource downloads consume their own CPU, RAM or bandwidth on whichever host runs them. |
| Network conditions | Latency, loss and retransmissions affect delivery and outgoing traffic. |
Load protection
Section titled “Load protection”Several implemented measures can help when load grows:
sync_interval = "auto"starts at 25 ms and can gradually increase to 45 ms under sustained load. Less frequent sync can reduce CPU work and traffic.[transport] bandwidth_budgetprovides an outgoing game-traffic target for automatic interval adjustment and the admission guard. A budget below the available uplink capacity leaves room for voice and other traffic.overload_guard = truecan pause new admissions if pressure persists. With automatic sync enabled, it waits until the interval reaches 45 ms. Admissions resume after recovery.
See load-protection messages for the reported load and state changes. The measurements above used a fixed 25 ms interval. For tests with automatic adjustment, time spent at each interval helps explain changes in the delivered update frequency.
Starting hardware allocations
Section titled “Starting hardware allocations”These suggested allocations include headroom. Actual requirements may be lower or higher, depending on a set of factors and your mod’s overhead.
| Connected players | CPU allocation | RAM | Available uplink |
|---|---|---|---|
| 100 | 4 fast physical cores | 16 GB | 1 Gbit/s |
| 300 | 6 fast physical cores | 16 GB | 1 Gbit/s |
| 500 | 8 fast physical cores | 32 GB | 1 Gbit/s |
| 1,000 | 12–16 fast physical cores | 32 GB | 2.5 Gbit/s port with at least 2 Gbit/s usable upstream |
These are starting points for host selection. CPU estimates assume dedicated time on cores with sustained performance comparable to the test host’s performance cores. RAM allocations allow for a real mod and service overhead. Databases, caches and resource hosting add their own requirements. A load test with your mod would help validate the estimates for 500 and 1,000 players.
SSD storage is recommended, with room for packages, scripts, persistent data, logs and a previous deployment for rollback. External hosting and a CDN can keep resource downloads off the game server’s uplink. Sustained outbound bandwidth and monthly transfer allowances are useful points to compare between hosting plans.
Runtime and OS requirements are listed in Requirements and compatibility.
Testing with your mod
Section titled “Testing with your mod”A staged load test can show how the host handles joins, leaves, streaming changes, vehicles, combat, database work and voice. Real GTA clients and representative network latency also help assess how the result feels to players.
The monitoring guide covers delivery, queue growth, script failures and disconnects. Recording the generators’ CPU usage and actual send frequency helps distinguish server limits from an overloaded test machine.
A useful starting value for max_players is the population that meets your
delivery target with room for workload spikes. Saving the tested configuration
and deployment versions alongside the results makes later comparisons easier.