-
Notifications
You must be signed in to change notification settings - Fork 6
Start How To Read This Wiki
What this page covers. The conventions this wiki uses to separate settled fact from active speculation, what each admonition box means, how claims were gathered and adjudicated, why no individual is ever named, and what the phrase "as of 2026-07-28" is doing at the bottom of pages.
Read this once. Everything else on the site assumes you have.
The headline: plain prose means the claim is confirmed. If a statement is not wrapped in a coloured box and does not carry an inline hedge, it has been checked against source code or against multiple independent measurements. Anything less certain than that is marked, and the marking tells you exactly how much less certain.
Unmarked prose is a confirmed fact. In practice that means one of the following was true during adjudication:
- The claim was settled from source code. The shipping
cmpunlockertree and 12 unreleased branch snapshots were read directly. Where documentation and code disagreed, the code won, every time, and the correction is stated in the text. - The claim was measured on hardware by two or more independent parties, with the measurements agreeing.
- The claim is arithmetic derived from values in the first two categories, with the derivation shown.
Some facts rest on a single report from a single machine. Those are never presented as settled, but they are also not hidden, because a single good measurement is far more useful than silence. They are hedged in the sentence itself:
"one tester reported ...", "observed once, on 2026-07-26, on one rig", "a single capture exists", "medium confidence".
When you see that phrasing, treat the number as a data point, not a specification. It has not been reproduced.
Four box types carry specific meanings. They are used sparingly and consistently.
Warning
Experimental
The material inside is on an unreleased branch, or rests on a single report, or both. It is not part of the shipping unlocker. PCIe Gen2, the 80 GB memory profile, the MIG enable and the multi-card handling all lived here at various times. Expect it to work on some hosts and not others, and expect the details to change.
Caution
Following the instructions inside can destroy hardware or lose data. This box is reserved for physical and irreversible risk: connector mis-mating, VBIOS flashing, thermal runaway, soldering, and allocations that fault the card. It is not used for "this might not work".
Note
Open problem
Nobody knows. The box states what is known, what has been tried, why the investigation stalled, and what one experiment would settle it. Every open item is collected and ranked by tractability on Open questions and tracked on the Status board. Thirteen of them can be settled by neither the documents nor the source trees.
If a number is not known, this wiki says it is not known. It never fills the gap with a plausible figure. A page that says "no CFM, static-pressure or fin-pitch figure for the stock heatsink exists in any source" is more useful than one that quotes an invented number, and the distinction matters more on this hardware than most, because several of the numbers in public circulation are wrong.
The underlying material is a mined archive of the community effort that produced the unlock:
project chat channels (clean-room and post-release), long-form write-ups, raw dmesg
captures, register dumps, screenshots, published tooling source, external references such as
teardowns and fuse tables, and the shipping and branch source trees themselves. Individual
claims in the working documents carry citation tags of the form
[artifacts__cleanroom_writeups#082] or [chat__unlocker__testing_general__01#039], plus a
confidence rating and the date the claim was established.
Raw claims were adjudicated into 24 domain documents, one per subject area (memory in four parts, firmware in four, compute in two, driver in two, PCIe in two, plus performance, thermal and power, VBIOS, tooling, provenance, troubleshooting, mods, NVLink, LLM inference and a miscellany). Each document separates canonical facts from dead ends, open questions, unresolved contradictions, corrections to the record, and a table of measured values.
Two of those sections deserve special mention because they are unusual and useful:
- Dead ends. Every failed hypothesis, with why it was plausible and what disproved it. Twenty-five in troubleshooting alone. These exist so the same wrong idea is not re-litigated. They surface on Dead ends.
- Corrections to the record. Places where project documentation, a public guide or a widely repeated claim is provably wrong against the source. Several of these are still circulating in third-party forks and AI-generated guides.
When sources conflict, this order decides:
| Rank | Source | Example |
|---|---|---|
| 1 | Shipping source code, read directly |
plmTable[] in 0001-sec2-postbl-plm-ss-cfg.patch
|
| 2 | Branch source snapshots |
0007-pcie-gen2.patch on Gen2
|
| 3 | Direct measurement with a posted capture | a verbatim dmesg block, a register readback |
| 4 | First-hand report without a capture | "one tester reported" |
| 5 | Project documentation |
README.md, docs/ARCHITECTURE.md
|
| 6 | Reasoned inference | explicitly labelled as such |
Project documentation ranks below measurement deliberately. The docs branch diverged
from the code: it invents acronym expansions, quotes the wrong SS0/SS1 values, references a
script that does not exist, and over-generalises the PLM readback rule. See the
Glossary.
A final pass compared every number that appears in more than one domain document against every other occurrence, re-deriving code-settleable disputes from the source trees rather than quoting the documents. It found 14 conflicts: 8 settled from code, 6 weighed on evidence. The result is a canonical value table that this wiki treats as authoritative. Where a page quotes a number that was disputed, it quotes the canonical value and, where the dispute is instructive, says what the disagreement was.
A worked example of why this matters: the archived 80 branch carries lmr: "0x0000028B" in
common/constants.yaml, and several documents repeat that as the value the branch programs.
It is not. build.sh never reads constants.yaml; the build sets 0x0000028A, the installer
prints 0x0000028A, and patch 0001 bakes 0x0000028A. Every tester who ran that branch
programmed a three-way-inconsistent geometry, which is itself the leading suspect for the
instability. Only reading the code exposes that.
No Discord handle, display name, real name or user ID appears anywhere on this site. Where attribution is load-bearing to the claim, this wiki writes "a researcher", "one tester", "two independent testers", "the maintainers", or "the tool author". Gendered pronouns are not used for anyone; they/them is used throughout.
There are three reasons, and only the first is about politeness:
- The work was done under clean-room rules. The governing standard was that no vendor secrets be discussed, that knowledge be admissible only where the same information is demonstrably derivable from public sources, and that posting leaked material was a bannable offence. An earlier server was destroyed because it may have contained leaked vendor material. Attribution invites exactly the kind of provenance argument those rules were designed to avoid.
- There is live legal exposure. A takedown notice was reported against at least one public fork on 2026-07-17, and the repository was observably offline. The report is first-hand but no takedown document is in the source set, so this one is medium confidence. Naming contributors adds risk to individuals without adding a single fact to the record.
- Attribution is not evidence. What matters for a technical claim is the capture, the register readback or the line of code, not who posted it. Removing names forces every claim to stand on its own basis, which is the correct outcome.
Community history is documented in Clean room and provenance without personalities. Prices, sellers, procurement and community disputes are out of scope entirely.
That is the adjudication date: the day the corpus was frozen, the source trees were read, and the cross-check was run. It is not the date the hardware was released, nor the date any individual claim was established.
Read it as "through the end of 2026-07-27". The capture happened just after midnight UTC, so 2026-07-28 itself contributes only three off-topic messages and no code. See methodology.
Three separate timelines run underneath it:
| Timeline | Range | What it covers |
|---|---|---|
| Hardware and board-repair record | from 2023-10-25 | Teardowns, die markings, power-rail repair, the A100 board comparison |
| VBIOS and fuse characterisation | from 2024-07, concentrated 2026-03 to 2026-05 | ROM diffs, fuse tables, cross-flash attempts |
| The unlock effort | 2026-06-27 to 2026-07-28 | Clean-room exploit work, then the shipping in-driver patch set |
Individual facts in the domain documents carry their own "established" date, and several carry a second date where they were corrected or code-confirmed during the 2026-07-28 pass. Where a date is significant to interpreting a claim, this wiki states it inline.
What the date implies for you, practically:
- The shipping unlocker as described here accepts exactly nvidia-open
610.43.03and610.43.02. If you are reading this after a newer driver has shipped, that whitelist has either been extended or the approach has moved. Check Driver versions. - Anything marked Experimental was still moving on the freeze date. The PCIe Gen2 hardcoded BDF bug, for instance, was root-caused on 2026-07-27 and fixed on one branch the same day, with several open reports never re-tested against the fix.
- The open-question list is a snapshot. Some of those items need one boot test to close.
- Nothing here has been re-verified against hardware after the freeze date.
Every substantive page follows the same shape:
- What this page covers, one short paragraph, plus the key result stated in the first two paragraphs. Nothing important is buried.
- Substance, densest first. Register maps and measured values are in tables. Commands, code and log output are in fenced blocks with a language tag, and are reproduced verbatim including addresses and error codes.
- Cross-links to the pages that go deeper. Links are relative and always end in
.md. - Open problems and superseded material, boxed.
Two structural conventions worth knowing:
- Speed and width are never conflated. PCIe generation (Gen1 to Gen2) is a driver-side software change that modifies no firmware image. Link width (x4 to x16) requires hand-soldering 24 missing 0402 capacitors. They are separate achievements with separate evidence, and no page mixes them. See PCIe Gen2 and Physical mods.
-
The two SKUs are never mixed. The 8 GB card (
10de:20c2) unlocks to 64 GB. The 10 GB card (10de:2082) unlocks to 40 GB. The 80 GB configuration for 10 GB cards was attempted and found unusable above roughly 40 GB. If a page seems to be talking about your card and quotes the other capacity, you are reading about the other SKU. Start at Identify your card.
Every claim on this site is traceable to a domain document, a citation tag and a date. The
most valuable corrections are the ones that come with a capture: a dmesg block, a register
readback with lspci -nn in the same paste, or a byte comparison. Several of the open
questions listed on Open questions would close with a
single such capture from a single card.
- Glossary for every acronym and term, including the ones commonly expanded wrongly.
- Risks for what can go wrong before you start.
- Methodology for the full adjudication procedure.
Start
Hardware
- Board And Variants
- Fuses And OTP
- GA100 Silicon
- Memory Subsystem
- NVLink Hardware
- Overview
- PCIe Subsystem
- Power Delivery
- Thermals
- VBIOS
Unlock
- Compute Throttle
- Driver Patches
- Falcon And Booter
- How It Works
- Memory Geometry
- Overview
- PCIe Gen2
- Privilege Level Masks
- Register Reference
- ROP Chain
Procedures
Operations
Frontier
History
Appendix