Game Server Benchmarks That Predict TPS: Single-Thread Share, Memory Bandwidth and Scheduler Latency

Game-server benchmark charts are usually built from workloads that look nothing like a tick loop. Cinebench saturates every thread, CrystalDiskMark measures sequential throughput, and network tests run on an idle link. None of those numbers predict the only thing players perceive: whether the simulation finishes each tick on time and whether autosaves cause a visible hitch. Four measurements do correlate, and each has a pass threshold you can check on a trial node within an hour.

Measurement 1: Single-Thread Share

Divide the single-core score by the multi-core score of the same machine. A 32-core EPYC 9354P scoring about 1,650 single and 48,000 multi has a share of 3.4%. An eight-core Ryzen 7 5800X3D scoring 1,520 single and 14,800 multi has a share of 10.3%. The second machine has a quarter of the cores and is far better for a simulation loop that lives on one or two threads.

CPU	Single-core	Multi-core	Share	Verdict for 60-tick hosting
EPYC 9354P 32c	1,650	48,000	3.4%	Poor per-instance, good for many small worlds
EPYC 4344P 8c	2,050	17,500	11.7%	Strong balance
Ryzen 7 5800X3D	1,520	14,800	10.3%	Large L3 cache helps modded worlds
Ryzen 9 7950X	2,180	41,000	5.3%	Best high-clock option, watch power limits
Xeon Silver 4310 12c	1,180	17,900	6.6%	Below threshold for 100+ players
Core i9-14900K	2,290	38,500	5.9%	Excellent clock, high power draw

Treat 7% as the floor for a 48-slot survival world and 9% for anything with heavy mod scripts. Re-measure after any hardware change — silent CPU substitution is the most common cause of a server that became noticeably slower without a config edit.

Measurement 2: Memory Bandwidth per Socket

Run a STREAM triad test on the populated node and require at least 45 GB/s on DDR4-3200 with eight channels and 70 GB/s on DDR5-4800 with twelve channels. A 128 GB allocation delivered as four 32 GB DIMMs in a twelve-channel board runs at roughly half the rated bandwidth, and the shortfall appears as 8–14% lower tick stability once AI populations exceed 300 entities.

For single-socket game nodes the practical check is simple: count populated channels. Four DIMMs on a consumer board with two channels is fine; four DIMMs on a server board with eight channels wastes three-quarters of the theoretical bandwidth.

Measurement 3: Scheduler Latency Under Fractional Load

An idle server has perfect tick times. What matters is behaviour when the box runs at 40–60% utilisation with a background backup competing for cores. Measure the 99th percentile of a synthetic 16.6 ms loop under that load. On a node with C-states enabled and a modest overcommit ratio, the 99th percentile lands at 18–26 ms — that is a tick overrun on 1% of ticks, which players feel as occasional stutter.

• Pass: 99th percentile tick duration under 17.5 ms at 50% load.
• Marginal: 17.5–19 ms; disable C-states and re-test.
• Fail: above 19 ms; the node is throttling or oversubscribed.
• Test with the same autosave and log volume production will generate, not on a bare install.

Measurement 4: NVMe Queue-Depth-1 Tail Latency

Sequential read of 5 GB/s is irrelevant to a game server. What matters is the time to complete a 4 KB write with one outstanding request, because that is what a save or a log flush looks like. Consumer drives deliver 60–110 µs at best and 400–900 µs at the 99.9th percentile. Enterprise drives with power-loss protection hold 25–60 µs with a 99.9th percentile under 250 µs.

Drive class	QD1 write mean	QD1 write p99.9	Save hitch on 2 GB world
Consumer NVMe (DRAM-less)	110 µs	900 µs	600–1,200 ms
Consumer NVMe (DRAM)	70 µs	450 µs	350–600 ms
Enterprise NVMe + PLP	35 µs	180 µs	120–220 ms
Enterprise NVMe RAID 1	40 µs	200 µs	110–200 ms

If the provider cannot tell you whether the drive has power-loss protection, assume it does not. Consumer drives on a 24/7 host also reach their endurance limit faster: a 600 TBW device under constant world saves and log writes can be exhausted inside two years.

Weighting the Four Into a Buying Score

A practical weighting for game hosting is 40% single-thread share, 25% scheduler latency, 20% storage tail latency, 15% memory bandwidth. Score each candidate from 0 to 10 against the thresholds above, multiply, and rank. A node that scores 9 on storage but 3 on scheduler latency will still produce stutter under load, and the weighted score reflects that ordering better than any single chart.

Run all four tests on a week-long trial before signing an annual contract. Providers that permit that testing and publish measured clock and NVMe details are the ones worth shortlisting — the provider profiles on the main site list those figures in a comparable format. For operators who would rather own the hardware, the same tests apply to colocation and bare metal deployments where you control BIOS, cooler and drive choice end to end.

The takeaway: a benchmark that saturates every core tells you how the machine behaves when doing nothing a game server does. Judge candidates on the four measurements that map directly to tick deadlines and save windows, and re-run them after every hardware change.