What 135 Measurements on a $200 Mini PC Actually Show

What 135 measurements on a $200 mini PC actually show is this: idle power between 12-14 W, a sustained 50 GiB write rate of 623 MiB/s on the internal NVMe, gigabit wire-speed networking, a 12.57-hour run with zero failures — and one USB port that is 8-10 times slower than the three next to it. The numbers below come from a single Beelink EQi12 (i3-1215U, 16 GB, 512 GB NVMe) measured in July 2026, with every value traceable to a published raw file.

This is not a review of one shopping choice and it is not a buyer’s guide. Both exist already: the full EQi12 review covers the purchase decision, and the mini PC home server roadmap covers planning. This page is the cross-dataset readout — what the measurement corpus as a whole says about what a budget mini PC can and cannot do as a home server.

Where the numbers come from

Every figure on this page comes from one of two places:

• The raw evidence files in the public repository eqi12-measurement-data, published under CC BY 4.0. Each table below names its source file.
• The individual lab pages on this site, which reproduce the same numbers with screenshots, methodology notes and their own honest-limit sections.

The machine under test: Beelink EQi12, Intel i3-1215U (6 cores / 8 threads), 16 GB DDR4-3200, WD PC SN540 512 GB NVMe, two Realtek gigabit ports, Intel AX200 Wi-Fi. Instruments: a wall-plug power meter, the same EAGET 1 TB USB SSD and cable for all port tests, and per-test logs with SHA256 verification for large transfers.

One boundary statement up front: all numbers are from this specific unit under Windows. Nothing here is extrapolated to other mini PCs, other CPUs, or Linux-specific performance. Where a number is derived arithmetic rather than a measurement, it is labeled as such.

Finding 1: the idle power profile is the whole running-cost story

Wall power was measured in seven states on 2026-07-13 with a power meter between the wall socket and the DC input (source file: EQi12_Power_EQi12_鍔熻€楁祴璇曡褰昣2026-07-13.txt, each reading photographed from the meter display).

Three things stand out. First, a busy four-container home server idles at 12 W — below the Windows desktop idling at 14 W. Containers doing nothing cost less than the desktop shell. Second, hardware transcoding is nearly free: 13 W while converting 4K60 to 1080p60 in real time, one watt above idle. Third, full load is bounded at 37 W — the entire worst case is a third of a typical desktop PC’s idle-plus-load spread.

The full state-by-state breakdown, including the invalid preliminary runs that were discarded, lives in the EQi12 power consumption lab page.

Derived arithmetic (not a measurement): at 12 W idle, the machine draws 0.288 kWh per day, about 105 kWh per year. At $0.15/kWh that is roughly $16 per year; at $0.30/kWh, about $32 per year. This is arithmetic applied to a measured wattage — your actual bill depends on your rate and uptime. The interactive version with your own rate and tariff inputs is in the home server power cost tool.

Finding 2: sustained stability matters more than peak benchmarks

The most operationally relevant dataset is not a benchmark at all — it is the monitored 12.57-hour run (source: EQi12_LongStability_12h57_Summary.txt plus 150-sample CSV and retained event logs).

The interesting number is not the zero failures — it is the maximum sampled CPU of 52%. A four-container stack (Jellyfin, PostgreSQL, nginx, Redis) running real home-server duty never pushed the i3-1215U past half its capacity. Combined with the memory floor of 5,472 MB free, the dataset says the 6-core/8-thread budget CPU was never the bottleneck in this workload class. The methodology and its caveats (a 5-minute sampling interval can miss sub-interval failures) are documented in the 12.57-hour stability test page.

Finding 3: internal storage is fine; the USB port lottery is not

The internal WD PC SN540 512 GB NVMe sustained a 623.29 MiB/s write and 468.2 MiB/s unbuffered readback across a complete 50 GiB transfer with matching SHA256 (source: EQi12_InternalSSD_50GiB_Result.txt). Sustained rate, not a cache-inflated peak — and comfortably above what a single gigabit client can pull over the network.

The external SSD tests produced the single most surprising result in the whole corpus. The same EAGET 1 TB SSD, the same cable, the same 50 GiB file, four different USB ports (source: EQi12_USB_鍥涙帴鍙?0GiB鏈€缁堝鐓?txt):

One port is 8-10 times slower than its neighbours, confirmed by complete transfers, not spot checks. There is no label on the outside of the machine that reveals this. If you attach a backup drive to the wrong port, your nightly backup window quietly becomes eight times longer. The follow-up analysis, including the warning against inferring USB generation from throughput alone, is in the USB storage page, and the troubleshooting path for slow ports is in the USB SSD limited to 40 MB/s fix page.

Finding 4: networking is exactly what the spec says

Directional 4 GiB transfers over each physical port measured 891-941 Mbps application-level throughput (sources: EQi12_Wired_NIC1_Direction1_Sender.txt, EQi12_Wired_Direction2_Receiver.txt; full methodology in the EQi12 network performance page). No surprise here — and that is the point. Two real gigabit NICs deliver real gigabit, which means the network will never be the reason a budget mini PC feels slow; storage layout and service configuration will.

Finding 5: media capability has a hard, measurable edge

The QSV dataset (sources: EQi12_QSV_4K60_to_1080p60_60s_Result.txt, QSV_2streams_Stream1/2.log, QSV_4streams_Stream1-4.log, codec matrix CSVs) draws a clean line:

• Two simultaneous 4K60 → 1080p60 H.264 transcodes run in real time on Intel QSV, at 13 W wall power.
• Four concurrent workloads do not complete in real time.
• AV1 decode works; AV1 encode is not supported by this generation’s QSV.

For a household of two to four viewers where at most two streams need transcoding at once, the machine is sufficient. For a media server shared with a larger group, or for an AV1 encoding pipeline, it is not. The complete codec matrix and the four-way failure evidence are in the Jellyfin QSV lab and the i3-1215U codec support page.

Finding 6: what the dataset does not cover

Honest limits, stated plainly:

• No temperature measurements. The dataset contains no thermal data in °C; thermal analysis was explicitly left out of scope.
• No Linux-native benchmarks. The benchmarks ran under Windows; a Proxmox or bare-metal Linux install may behave differently.
• Component-level only. The nginx/pgbench/Redis numbers measure the container stack on this machine — they are not a substitute for testing your specific application.
• One unit. Two EQi12 units exist in the project, but most datasets come from unit A. Unit-to-unit variance is unmeasured.

The application-level benchmark detail (including the failed preliminary runs that were discarded and why) is in the Docker workload benchmark page.

The citable fact block

If you quote this page, quote this block — it carries its own provenance:

The whole corpus — 135 measurements across power, storage, network, stability, codecs and security — is available under CC BY 4.0 at github.com/sohasteve-pixel/eqi12-measurement-data.

What this means if you are shopping at this price

The dataset supports a specific, narrow claim: a $200-class 6-core mini PC is a competent low-power home server for container workloads, light databases and two-stream media transcoding — and its real-world risks are configuration-level (a slow USB port, sleep-state wake quirks), not performance-level. The failure modes worth planning for are the boring ones: power-loss recovery after outages, backup restores, and port selection — each of which has its own evidence-backed page on this site.

If you are choosing between machines rather than evaluating one, start with the mini PC home server buying guide and the Windows vs Proxmox on the same mini PC comparison; both link back into the lab pages that carry the raw numbers.