Squad fills 100-slot servers with players who will absolutely notice a dropped tick. A full 100-player match spawns dozens of vehicles, hundreds of deployables and a single-threaded game loop that already runs near the CPU’s limit on stock settings. Adding cores does almost nothing past a point; what matters is which cores the server thread lands on, how the OS schedules the rest, and whether memory bandwidth keeps up.
How Squad uses CPU at full population
Unreal Engine 4 dispatches most gameplay logic on the game thread, with physics and rendering work split across a handful of helper threads. Squad’s server build is lighter on rendering but heavier on replication: every player position, inventory change and vehicle state must be serialised and multicast. At 100 players the replication thread becomes a measurable cost alongside the game thread.
Typical utilisation on a modern 8-core desktop part sits around 45 to 65%, but the peak is concentrated on two or three cores while the rest idle near 20%. That imbalance is why a 6-core chip at 4.8 GHz can feel smoother than a 16-core server part at 2.6 GHz. The bottleneck is single-thread speed first, core-to-core penalty second, memory third.
What thread affinity actually changes
Affinity pins a process or thread to a defined set of CPUs with taskset or numactl. On a single 100-player instance, pinning the whole process to four adjacent physical cores forces the game thread, replication and networking onto the same CCX, which keeps their shared cache hot and removes cross-socket trips. Latency of a cache miss local to a CCX is roughly 40 to 50 cycles versus 100 or more across sockets.
On dual-socket hardware the effect is larger. Without affinity, the scheduler may migrate the replication thread to a different socket mid-round, producing occasional 5 to 10 ms frame-time spikes. Pinning each instance to its own socket or NUMA node removes that migration cost. Measure with PerfMon or perf rather than assuming, because the benefit ranges from negligible to 10% depending on the board.
Where affinity does not scale
Pinning two full instances to the same four cores is a mistake. Squad’s thread pool grows under load, and when two game threads compete for the same physical core, each loses the shared 256 KB L2 and the context switches add jitter. On a 12-core machine the sane split is one instance per socket, or two instances on six cores each with hyperthreading disabled for the second set.
Beyond that, affinity cannot fix a CPU that is simply too slow. A 100-player Squad round on a 2.6 GHz server Xeon will hit 80% game-thread utilisation and drop tickrate regardless of pinning. Move to a 3.5 GHz+ part before tuning threads, and expect the gain from affinity alone to stay in the 3 to 8% range once the CPU is adequate.
Frequency, cache and memory bandwidth
Single-core turbo frequency is the dominant factor. A Ryzen 9 7900X running its game thread at 5.4 GHz handles a 100-player round with headroom, while a Xeon Silver 4314 at 3.4 GHz maximum sits close to the limit. ECC memory and server boards matter for reliability, but they are not the bottleneck; core speed is.
Memory bandwidth starts to matter around the 80-player mark when vehicle convoys and artillery effects spike replication volume. DDR5-5600 dual-channel delivers close to 90 GB/s, well clear of what Squad needs; a quad-channel server platform adds headroom but costs more per core. Cache size helps more: 32 MB or more of unified L3 keeps the map and player lists resident between ticks.
Practical tuning recipe
Start with the BIOS: disable power saving, fix the all-core frequency above 4 GHz, and turn SMT off for game instances if you run two per server. SMT helps brute-force workloads and often hurts latency-bound simulation because threads share execution units. Then set the process affinity and verify with htop that no single core is pinned permanently at 100%.
In the server config, lower MaxTickRate to 30 if your community tolerates it; 60 ticks at 100 players is a heavy replication load and most players accept 30 for admins and vehicles. Reduce the number of tracked deployables and cap vehicle counts where the map allows. These settings often save more CPU than any scheduling tweak.
Measuring before you buy
Use PerfMon’s processor time per logical CPU, plus a frame-time logger for the game thread, and collect data on a full 100-player round rather than an empty server. Look for a single core sitting above 85% and a tick-rate counter slipping below target. That pattern tells you the game thread is the wall, and more cores will not help.
If two cores sit near 100% and the rest idle, move to a higher-clock CPU instead of adding cores. If the load is spread but the tickrate still drops, suspect memory latency or network. Our dedicated server comparison table lists per-core frequencies for game-oriented hosts, which is the fastest way to rule out a slow part.
Scaling beyond one instance
- One 100-player instance maps well to four physical cores on a 4.5 GHz+ chip.
- Two instances need separate CCX or socket affinity, not just core count.
- Disable SMT for latency-bound instances.
- Keep the game thread on a single fast core; do not split it.
- Buy frequency first, then cores, then bandwidth.
Thread affinity is a tuning tool, not a scaling strategy. It recovers 3 to 10% when applied to a fast enough CPU and costs you stability when applied to a slow one. Size the hardware around single-core clock speed and replication load, pin only after you have measured, and the dedicated hosting fundamentals guide covers the rest of the server build.

Leave a Reply
You must be logged in to post a comment.