Open first-party hardware benchmarks and reusable specification datasets, with raw files, methods, licenses, and correction history.
59 published measurement rows covering five 3DMark workloads, clocks, thermals, board power, DLSS 4 Multi-Frame Generation, PC latency, and native 4K gaming.
The raw file, the method behind it, and the page or tool that uses it sit together.
126 current components across seven categories with sockets, chipsets, RAM support, dimensions, power requirements, source pages, and per-row verification dates.
Card-specific tune points with measured power, thermal, and FPS changes. Every populated number cites where it came from; blank measurements stay blank.
Decode, prefill, VRAM, and load time for Ollama and llama.cpp on the same RTX 5080 across three models, wall-clock cross-checked with llama-bench as a third instrument. CC BY 4.0.
Measured resident memory, startup time, and tool counts for 11 popular MCP servers plus a concurrent five-server stack, with per-iteration arrays and a two-path cross-check. CC BY 4.0.
Dated $/GB index for mainstream 32GB (2x16GB) DDR5 kits at named direct US retailers, captured with a published method during the 2026 memory price crisis. CC BY 4.0.
Dated street price versus verified launch MSRP for eight RTX 50 and RX 9000 graphics cards at named direct US retailers, with a published capture method. CC BY 4.0.
Dated $/TB index for twelve 2TB NVMe drives that Newegg sells itself, with a published capture method. CC BY 4.0.
Dated $/TB index for 8TB 3.5-inch NAS, surveillance, enterprise, and desktop hard drives at named direct US retailers, captured with a published method during the 2026 drive supply squeeze. CC BY 4.0.
Measured board power, tokens per joule, and electricity cost per million tokens for local LLM inference on a retail RTX 5080 — 1Hz telemetry paired with decode rates, four models from 3B to 20.9B. CC BY 4.0.
Measured decode and prefill tokens/second, load times, and VRAM residency for local LLMs on a retail RTX 5080, with full reproducibility metadata. CC BY 4.0.
Open specifications and compatibility data for CPUs, GPUs, motherboards, memory, storage, power supplies and cases. Every row carries a last-verified date; source URLs are included where available. CC BY 4.0.
First-party 3DMark, clocks, thermals, board power and 4K gaming results measured on the operator's own card.
Open, crowdsourced GPU undervolt settings with measured power, thermal and FPS deltas, each row citing a resolvable source.
Owner-reported coil whine for GPUs, PSUs, and AIO pumps on an anchored 0-4 severity scale, published as per-model distributions (n and share at each severity) once a model reaches 5 reports — odds, not verdicts. CC BY 4.0.
Owner-measured graphics-card droop in millimetres, one row per card with length, mounting, support, months installed and slot type, published as per-model distributions once a model has 5 unsupported horizontal readings. CC BY 4.0.
82 in-stock RTX 5090 prebuilt listings from the 2026-08-17 Newegg feed: stated PSU wattage, efficiency rating and 16-pin cable disclosure per listing. Open CSV, CC BY 4.0.
Owner readings of the current on each of the six 12 V pins of a 16-pin graphics card connector, grouped by cable type. A group publishes at five readings taken under load. CC BY 4.0.
What owners of OLED monitors see on their own panels, on a fixed scale from 0 to 4, with months in use, panel hours and warranty outcomes. A panel type publishes at five monitors. CC BY 4.0.
The time a PC's firmware takes before the system starts to load, as each machine recorded it for its last boot, with the board, BIOS and memory beside it. A group publishes at five readings. CC BY 4.0.
What hard drives report on the day they arrive: SMART hours, Seagate FARM hours and dead-on-arrival counts, by seller and listing condition. A group publishes at five drives. CC BY 4.0.
Memory in use on Windows PCs that have just restarted, read from Windows' own counters, by installed memory. A group publishes at five machines. CC BY 4.0.
Core and hotspot temperatures from the same sustained-load session, with tool versions, exact-model bands at five qualifying rows and a GPU-family rollup at twenty. CC BY 4.0.
TechFuelHQ publishes first-party benchmark datasets and reference compilations. The cards above link to each dataset’s method, dates, attribution and correction history. Check those notes before comparing measurements or reusing a file; price observations describe their capture dates, not today’s prices.
Eight of the datasets are censuses that take readings from owners: coil whine, GPU sag, 12V-2x6 pin current, OLED burn-in, POST time, drive arrival, Windows idle memory and GPU hotspot delta. Each opened empty. Every primary group waits for five counted rows; the hotspot census also waits for twenty before publishing its family rollup.
Start with the question you want to answer. These three packages contain our published RTX 5080 measurements, with the test conditions on each dataset page. The row counts below describe the downloadable files, not separate test systems.
| Your question | Dataset to use | Keep these conditions beside your chart |
|---|---|---|
| What changed between the Astral’s stock and overclocked profiles? | Astral RTX 5080 benchmark package: 59 data rows across the published charts | Match the chart, workload, tune and metric. Generated frames and native rendered frames answer different questions. |
| How did model, context and prompt length affect local inference? | Local LLM throughput: 34 data rows | Keep the model digest, quantization, context, prompt length, Ollama version and capture date. Decode and prefill are separate measurements. |
| How much GPU energy did these four inference workloads use? | LLM power efficiency: four model rows | These are GPU board-power measurements, not whole-PC wall readings. Retain the test settings and the stated electricity rate when quoting costs. |
For your own spreadsheet or article, download the benchmark reuse index (CSV). It includes the dataset page, raw CSV, evidence manifest, license and a suggested credit line for each package. Link to the dataset page when reusing a result so readers can inspect the method and any later corrections. All three packages use CC BY 4.0; keep attribution and identify any changes you make.
These links open raw files. CSV files work in a spreadsheet; TXT and Markdown files contain evidence notes. Read a Python script before running it, and follow the method and license on its accompanying page.
The published datasets use Creative Commons Attribution 4.0 unless a dataset page states otherwise. The exact attribution line and license file travel with each dataset.
Questions, requested fields, or a challenged measurement: contact@techfuelhq.com.
