Measurements

What our VPS actually do

All four plans put through the bench, each reading with its method. The figures that do not flatter us are here too: a chosen measurement is no longer a measurement.

Method · Every reading was taken on production VPS identical to the ones we sell, ordered through the public checkout, with no special tuning. Each table states the tool and its parameters. Readings taken on 12 and 13 November 2026.

Advertised against measured

The first table is the one that matters most: it puts what the price list promises against what the machine delivers, plan by plan.

What the storefront promises, checked on every plan

Taken on four VPS ordered separately, at delivery.

PlanvCPU and memoryDiskThroughputGeekbench 6
Start · 2 vCPU2 vCPU · 4 GB40 GB real / 40 advertised959 / 1,000 Mbit/s2,272 measured
Medium · 4 vCPU4 vCPU · 8 GB80 GB real / 80 advertised961 / 1,000 Mbit/s3,879 measured
Pro · 6 vCPU6 vCPU · 16 GB120 GB real / 120 advertised963 / 1,000 Mbit/s4,789 measured
Ultra · 8 vCPU8 vCPU · 32 GB160 GB real / 160 advertised959 / 1,000 Mbit/s5,618 measured

Processor

Two Intel Xeon Gold 6238 per hypervisor. The scores below were taken on our own VPS and published: every row links to its result, which anyone can open.

Geekbench 6

Geekbench 6.5, one run per plan, results published on browser.geekbench.com.

PlanSingle-coreMulti-corePublic result
Start · 2 vCPU1,2752,27219461712 ↗
Medium · 4 vCPU1,2593,87919462046 ↗
Pro · 6 vCPU1,2504,78919462136 ↗
Ultra · 8 vCPU1,2325,61819462240 ↗

Sustained load and neighbours

sysbench, 100% of the vCPU for 5 minutes, then all four VPS on one hypervisor started together.

MeasurementResult
Throughput held after 5 minutes at full load97%
Throughput kept with all four VPS flat out89 to 99%
Steal time taken by the hypervisor0.02%

Disk

Four 12 Gb/s SAS SSDs in RAID 1+0 behind an HPE Smart Array controller. Direct I/O bypasses the cache: without it, these figures would be measuring memory.

4K random and sequential

fio, direct I/O, 128 parallel requests. 1 MB blocks for the sequential run.

Plan4K read4K writeSeq. readSeq. write
Start · 2 vCPU172,701 IOPS165,714 IOPS3,787 MB/s2,161 MB/s
Medium · 4 vCPU163,205 IOPS164,824 IOPS3,783 MB/s2,156 MB/s
Pro · 6 vCPU162,073 IOPS158,873 IOPS3,758 MB/s2,161 MB/s
Ultra · 8 vCPU164,409 IOPS163,297 IOPS3,778 MB/s2,159 MB/s

Latency under mixed load

Mixed load, 70% read / 30% write, 64 parallel requests.

PlanMedian latencyp99 latencyMixed 70/30
Start · 2 vCPU305 µs408 µs142,994 IOPS
Medium · 4 vCPU354 µs444 µs140,627 IOPS
Pro · 6 vCPU354 µs424 µs138,140 IOPS
Ultra · 8 vCPU375 µs449 µs140,670 IOPS

Steadiness over 24 hours

Taken every hour for 24 hours on the Start plan, peak hours included.

MeasurementMedianPeak 18:00-23:00Night 02:00-06:00
Processor (sysbench, events/s)1,9321,9371,944
Disk 4K (IOPS)149,110147,122150,122
4K latency339 µs341 µs344 µs

Network

The link is advertised at a guaranteed 1 Gbit/s, with access to the hypervisor's 10 GbE above that. For a French audience the question that matters is not throughput, it is latency from a home connection.

From consumer home connections

Globalping: 17 home connections across France, 16 pings an hour for 24 hours. The reference column measures the same connections against a third-party server in Paris, which removes the subscriber's own line and leaves only the path to us.

OperatorLinesTo the VPSTo the referenceDifference
Bouygues Telecom33.9 ms3.3 ms+0.8 ms
Free54.4 ms3.1 ms+1.4 ms
SFR49.8 ms10.7 ms−0.3 ms
Orange512.2 ms10.9 ms+1.3 ms

Round trip from the VPS, over 24 hours

Median over 24 consecutive hours, peak hours included.

DestinationMedianp99p99.9
Paris1.2 ms1.4 ms1.7 ms
London8.2 ms9.0 ms12.3 ms
Amsterdam9.9 ms10.3 ms10.5 ms
Frankfurt10.7 ms11.1 ms11.3 ms
New York78.9 ms79.6 ms83.7 ms

International throughput

iperf3 against the public servers used by the YABS script, inbound then outbound.

DestinationInboundOutbound
London953 Mbit/s1,120 Mbit/s
Amsterdam937 Mbit/s1,120 Mbit/s
New York824 Mbit/s1,050 Mbit/s
Singapore720 Mbit/s885 Mbit/s

By use case

The same machines put to work on what people actually run. Each row compares the entry plan with the largest one.

From Start to Ultra

Each use case with its real software and real load: WordPress with WooCommerce and no page cache, PostgreSQL 17 with 16 clients, llama.cpp with no graphics card, Paper 26.3 for Minecraft, WireGuard between two VPS, the official Nextcloud 35 image.

Use caseMeasurementStartUltra
WooCommerce shop, no cachedynamic pages/s21107
PostgreSQL 17, 16 clientstransactions/s3,6948,611
PostgreSQL 17, read-onlytransactions/s24,624100,469
MariaDB 11.8, 16 connectionstransactions/s1,2053,723
C++ build (llama.cpp)duration555 s183 s
Next.js buildduration11 s11 s
WireGuard tunnelthroughput935 Mbit/s935 Mbit/s
Nextcloud, 1 GB filedownload84 MB/s85 MB/s
Minecraft Paper, startupduration17.1 s14.8 s
Qwen3 4B, no graphics cardtokens/s4.212.1
720p live transcodingsimultaneous streams2.66.8
Virtual machine inside the VPSprocessor kept97%97%

Hosting

Navigation