METHOD / 2026-10-24 22:15 UTC
Evidence before a verdict.
This is a small independent English index of seven public GPU unlock reports. It is not an exhaustive hardware database, an installation guide, a compatibility certificate or an official NVIDIA resource.
Why this page exists
Public reports use different GPUs, operating systems, power limits and workloads. This index brings exact identities, original links and comparison limits into one filterable view. The upstream project already has a useful tested-cards table; read it as well. Newer reports can appear after either summary was last updated.
How the labels work
- Paired GEMM
- The author supplies locked and unlocked dense-matrix measurements. We list important control differences. This does not mean independent reproduction or freedom from confounding.
- No locked baseline
- Only unlocked measurements, or comparisons between different unlocked settings. We do not infer an unlock uplift.
- No BF16 uplift reported
- The author reports no double-rate behavior for the stated board and test. It does not prove zero effect for every workload or explain the cause.
- No matching report
- No result matches your filters within these seven reviewed entries. It says nothing conclusive about reports elsewhere or compatibility.
Sources and snapshot boundaries
Each entry links to its original author. The dates on cards are report dates, not dates of our hardware tests. We have not tested this hardware. Issue discussions are live and may change after our review.
- RTX 3060 12 GB / Linux — Santha-SK
- RTX 3080 12 GB / Linux — trevelyen
- RTX 3070 8 GB / Linux — sakusaku0214
- RTX 3090 24 GB / Linux — chflame163
- RTX 3080 Ti 20 GB OEM / Linux — 3615074213
- RTX 3090 24 GB / Unraid — ampersandru
- RTX 3090 FE 24 GB / Windows — impacifist
Additional primary references: Linux project README at the checked revision; Windows compatibility notes at the checked revision.
Raw TFLOPS from different cards, libraries, matrix sizes and power caps are not ranked or averaged. A model name is not used to fill in a missing PCI ID. The Unraid report therefore keeps its ID unknown. Microbenchmark gains are not converted into application claims.
The time calculation
For a fixed amount of work, let g be the throughput increase as a fraction and p the fraction of the original runtime that benefits. The remaining time fraction is (1 − p) + p ÷ (1 + g). At g = 0.8 and p = 1, the time saving is 44.4%; at p = 0.5 it is 22.2%. This simplified scenario omits changes in overhead and bottlenecks.
Corrections and review
This site was assembled with AI assistance and source review. Original reports remain third-party claims. For newly published information, follow the source links rather than treating this snapshot as live. For an index correction, contact the index maintainer on X with the entry and a public source. Do not send private system logs, identifiers or credentials.
Privacy and funding
No account, upload or hardware access is needed. Filters and the calculator run in your browser. Filter selections appear in the page URL so a view can be shared. This site has no advertising. Optional Cloudflare Web Analytics loads only after you choose Allow measurement, on the custom domain. It measures page visits and performance; we do not send custom filter or calculator events. Your choice is stored in this browser, and you can change it using Measurement choices in the footer. Browser Do Not Track and Global Privacy Control signals disable this optional script. Self-test URLs do not load it. Optional measurement can undercount visitors. The hosting provider may process ordinary server request data. See Cloudflare data collection details. Source links take you to GitHub, which has its own privacy policy. No affiliation or endorsement is implied.