BitCloudVPS benchmarks - Geekbench, fio, iperf by location

Published performance snapshot for BitCloudVPS locations: Geekbench-class CPU figures, fio NVMe, and iperf bandwidth notes across all twenty-one datacenters.

Published 2026-08-30 · Updated 2026-10-21 · BitCloudVPS Editorial

Method

Representative production-class readings for the Nano/Starter SKU family (EPYC vCPU, ECC, local NVMe, 1 Gbps port). Re-run on your instance after deploy. Last refreshed in October 2026.

CPU (Geekbench 6 class)

Disk (fio) and network (iperf)

Deploy Nano from $3 and verify on your node. Token-only signup available.

Reading the numbers: what a shared NVMe host should show

Benchmark results only mean something against the hardware they run on. Our Starter tier pairs 2 vCPU of an EPYC 7443P with 4 GB ECC RAM and 60 GB NVMe, so a fair expectation is Geekbench-6-class single-core scores typical of that Zen 3 chip — not the inflated numbers dedicated Xeon E-2388G or dual-EPYC 9474F boxes (our Start $79.50 to Dedicated Ultra $449.50 dedicated line) produce.

Disk is the more common place to hide overselling. A healthy NVMe share should sustain random 4K writes without collapsing to hard-drive speeds under fio's 4k randwrite test, and the 10 Gbps port ceiling should show up in iperf3 to well-peered targets — we run 21 sites with peering at exchanges like AMS-IX, DE-CIX, SGIX, and HKIX.

Reproducing these tests on your own server

Every number on this page is reproducible with public tools, which is the point. Install fio and iperf3, run a 30-second 4k random-write test against your data volume, and a 10-stream iperf3 pull from a nearby exchange-connected host. Geekbench 6 runs unmodified on any Linux VPS. If your workload is latency-sensitive, also ping a target in the same city: same-city RTT against racks in Tokyo or Singapore should sit in low single-digit milliseconds.

Run the same suite before and after a plan change (Nano $5.49/mo up to Max $60/mo) — the per-tier CPU and disk allocation is what changes, and the deltas should be visible in exactly these three tools.

Why the same plan benches differently across 21 sites

Two servers on the same plan and the same CPU will not post identical numbers in Reykjavik and Bangkok, and the difference is mostly network, not silicon. Each of the 21 sites carries its own uplink tier and its own peering mix: Zurich, Frankfurt, Tokyo, Singapore, Ashburn and Dallas sit on 100 Gbps uplinks; Amsterdam, Hong Kong, São Paulo, Latvia and most others on 40 Gbps; Reykjavik, Romania and Bucharest-class sites on 10 Gbps. Peering decides how far your packets travel to reach the open internet — an Amsterdam box plugs into AMS-IX and DE-CIX, Frankfurt into DE-CIX Frankfurt, Zurich into SwissIX, and a Bangkok box into BKNIX — so iperf results depend on where you test from as much as where the server sits.

Read the two layers separately. The port speed on every VPS plan is an unmetered 1–10 Gbps, which caps what a single stream can show; the site uplink caps what a thousand streams can show. CPU and disk, by contrast, are rack-independent at the plan level: every plan runs NVMe with ECC RAM, and Starter's 2 vCPU come from an EPYC 7443P at every location — a Geekbench run in Almaty and one in Zurich measure the same silicon behind different shared tenancy.

The practical rule: benchmark the CPU and disk to compare plans, and benchmark the network to compare locations. A disk score that differs 15% between two sites is tenancy noise; a network score that differs 10x between a 10 Gbps and a 100 Gbps site is the uplink doing exactly what the catalog says it does.

Single-core, all-core, and what an EPYC 7443P class host means for your scores

Geekbench-style CPU results on shared VPS plans are dominated by two numbers that mean different things. The single-core score reflects the clock behavior of one thread — what a web request, a language interpreter or a build step actually feels like. The all-core score reflects how many of your allocated vCPUs the scheduler lets you saturate and for how long. On a shared host these diverge under load: a quiet neighbor leaves your all-core run at full allocation; a busy one shows up first in sustained all-core runs, rarely in single-thread bursts.

That is the honest reason we describe the platform as EPYC 7443P-class per allocation rather than printing one guaranteed score per plan. Starter carries a documented 2 vCPU on that platform with 4 GB ECC, and the same silicon family backs the rest of the VPS line; what varies run to run is tenancy, not the model. When you compare our numbers against any other shared host, ask three questions: single-core or all-core, how many runs were averaged, and whether the run happened on an empty or a warmed machine. A single all-core run on a fresh VM tells you almost nothing reproducible.

For capacity planning, weight your own workload before the benchmark. If your process is single-threaded — most web apps behind a fast proxy are — the single-core number predicts your latency. If you batch, compile or render, provision for the all-core allocation and test with your real job for an hour; a synthetic 60-second stress is a weaker instrument than your own build pipeline, and it is free to run on the plan you are evaluating.

The plan-to-benchmark map: ports, storage and what each tier can show

Benchmarks only mean something against the spec they run on, so here is the ceiling each tier carries. The constants first: every VPS plan includes NVMe storage with ECC RAM, an unmetered 1–10 Gbps port, 1x IPv4 and a /64 IPv6 block, at the same price in all 21 locations. The variables are the allocations:

How to benchmark against it: run fio with direct I/O against the NVMe volume and expect variance from shared tenancy — the honest comparison is your own repeated runs on the same plan, not our numbers versus a dedicated server's. Run iperf and expect the unmetered 1–10 Gbps port to be the cap on a single box, with the site uplink (40 Gbps at most sites, 100 Gbps at Zurich, Frankfurt, Tokyo, Singapore, Ashburn and Dallas) as the ceiling for aggregate load.

If your fio results outgrow the whole VPS line, that is the intended signal to move to the storage SKU (1–5 TB NVMe, $24–$56/mo, up to 8 vCPU / 32 GB at the top) or to dedicated metal with IPMI — same checkout, same coin list, no new KYC step.

Six mistakes that make a VPS benchmark lie to you

Most "my VPS is slow" reports are measurement artifacts. These are the six that account for nearly all of them. One: running fio against the filesystem cache — without direct I/O flags you measure RAM, not the NVMe, and scores look impossibly good. Two: single-stream iperf on a host with real RTT — one TCP stream cannot fill a 1–10 Gbps port over distance, so the test underreports the port; use parallel streams and a nearby endpoint, then a far one, and label both. Three: benchmarking during your own background load — a backup, an IBD sync or a build running concurrently will halve disk scores that are otherwise stable.

Four: one run, no repetition. Shared tenancy means any single run is a sample; take three runs minutes apart and read the median, and treat a 10–20% spread between runs as normal rather than as degradation. Five: comparing across different locations and blaming the plan — a 10 Gbps-uplink site will legitimately trail a 100 Gbps site on network tests while posting identical CPU scores; isolate the layer you are judging. Six: trusting the first boot — a fresh VM may still be settling (provisioning tasks, package updates); let it sit ten minutes before the first serious run.

What remains after subtracting the artifacts is small: real plan differences show in allocated vCPU count and RAM, real location differences show in uplink and peering, and everything else is noise you introduced. The reproduction commands live in the section above — run them yourself on the cheapest plan that matches your workload, because a benchmark you ran is worth more than one you were shown.

What to expect from the plan you ordered, tier by tier

Every VPS plan runs on NVMe with ECC RAM behind an unmetered 1-10 Gbps port, so the floor is the same from Nano to Max: no spinning disks, no oversold bursts pretending to be baseline. What changes between tiers is headroom. Starter (2 vCPU EPYC 7443P, 4 GB ECC, 60 GB NVMe) is where most single-service benchmarks land; Business and above add cores and NVMe depth that show up in all-core fio and compile-style runs, not in single-thread scores.

If your disk numbers are well below the NVMe class you expect, check three things before blaming the rack: whether you are benchmarking the shared queue at peak hour, whether your test file is small enough to sit in host cache, and whether the port speed on your plan is shaping the iperf leg of the run. A 1 Gbps plan cannot show 10 Gbps fabric, and that is a plan property, not a defect.

Where the limits are (October 2026)

Keep the threat model precise. Anonymous-adjacent hosting removes the host-side identity trail: token-only signup, optional email, no KYC review, crypto invoices that record a txid rather than a name. It does not blind the host to traffic it must carry, it does not make a logged-in panel session anonymous, and it does not protect workloads that identify themselves — an admin login from your home IP defeats the whole design. Exact limits of no-KYC anonymity. OPSEC guide. Full checklist. Payment-side note: chain analysis happens on public blockchains, not at the host. Coin choice and network choice are your OPSEC knobs, and the checkout keeps 19 of them open.