实测,而非宣称
价格背后的数字
下列每个数字都来自我们自己的机群遥测,每小时采样一次,取最近 90 天平均值。原始导出数据放在运行状态页。
- 单线程吞吐
- 2,240
- 随机读 IOPS
- 512,000
- CPU steal
- 1.4 %
- 24 个月可用率
- 99.987 %
01
单线程吞吐
单线程吞吐
Geekbench 6 单核,独享核心实例,机群中位数
Geekbench 6 — single core
Echelon VPS — EPYC 9354P (Genoa)2,240
EPYC 7443P (Milan)1 674
Xeon Gold 62481 246
Xeon E5-2680 v41 021
Geekbench 6 — multi core, by plan
Nano — 1 vCPU2,180
Micro — 2 vCPU4,210
Base — 4 vCPU8,060
Work — 6 vCPU11,740
Forge — 8 vCPU15,380
Anvil — 12 vCPU21,620
Monolith — 16 vCPU27,940
02
存储
存储
fio 4K 随机读,队列深度 32,在有负载的宿主机上
Nano — 25 GB96,000 IOPS
Micro — 50 GB128,000 IOPS
Base — 100 GB186,000 IOPS
Work — 200 GB224,000 IOPS
Forge — 400 GB312,000 IOPS
Anvil — 800 GB428,000 IOPS
Monolith — 1600 GB512,000 IOPS
4K read latency
0.08 ms
Sequential read
6.8 GB/s
Sequential write
4.1 GB/s
Drive endurance
1.3 DWPD
03
CPU steal
CPU steal
全机群第 95 百分位。共享套餐超售上限为 3:1,独享核心套餐为 1:1。
独享核心0.0 %
共享核心1.4 %
Industry budget tier11.8 %
- cpu
- AMD EPYC 9354P — Genoa, Zen 4, 3.25 GHz base / 3.80 GHz boost
- memory
- 384 GB DDR5-4800 ECC registered per host
- storage
- Samsung PM9A3 Gen4 NVMe, RAID-10, hot spare per chassis
- network
- Dual 25 GbE to separate leaf switches, LACP
- power
- A+B feeds, N+1 UPS, on-site diesel with 72 h fuel
- hypervisor
- KVM on Linux 6.12 LTS, nested virtualisation enabled
- oversubscription
- 3:1 on fair-share plans, 1:1 on dedicated-core plans
04
自己跑一遍
自己跑一遍
随便开通什么,用你信得过的测试套件跑一遍。如果数字令你失望,72 小时内申请退款。无需填表,也无需说明理由。
01How can it be this cheap and still be fast?
Three reasons, none of them exciting. We buy last-generation-plus-one EPYC hosts outright instead of renting capacity from a hyperscaler. We sell terms, so a host is full the day it is racked instead of three months later. And we do not employ a sales team. What we do not do is oversubscribe past 3:1 on fair-share plans or past 1:1 on dedicated-core plans — the numbers on the performance page come from production hosts under real load.
02Fair-share or dedicated cores — which do I need?
Fair-share is right for almost everything: web apps, databases under a few hundred queries a second, game servers, VPN endpoints, CI that is not on the critical path. Choose dedicated cores when a workload is CPU-bound for minutes at a time — video transcoding, compilation, backtesting, inference — or when a scheduler's p99 matters more than its median. It adds 55% to the plan price and removes neighbours from your cores entirely.
03What CPU steal should I expect?
Fleet-wide, the 95th percentile on fair-share plans sits at 1.4% and on dedicated-core plans at 0.0%. We publish the rolling figure on the status page. If a host drifts above 3% steal we stop selling on it and migrate instances off, which is why some sites show 'low stock' from time to time.
04Are backups included?
Two manual snapshots per instance are free and always available. Scheduled backups are an option priced as 10% of the compute line for weekly and 20% for daily — weekly keeps four restore points in the same country, daily keeps fourteen in a second one. Restores are self-service from the console and take minutes, not a ticket.
05Can I resize a server mid-term?
Upwards, at any time, with the difference pro-rated across the remaining term. Downwards at the end of the term. Memory and vCPU changes need a reboot; disk grows online. Moving between regions is a rebuild, not a resize — the data does not follow.
06Do you offer Windows?
Windows Server 2019, 2022 and 2025 Standard, with the SPLA licence included in the monthly figure shown at configuration. Remote Desktop is open on first boot and the licence is ours, so you do not need to bring one.