Refer to Gen2 as what the unlocker does, not as a branch
Gen2 shipped in master on 2026-07-29, so anyone who installs the unlock gets
Gen2 and only a card with no unlock reports Gen1. The state tables readers use
to identify their card still said 'after the Gen2 branch'.
Rewrites those columns across the physical mods, PCIe subsystem, overview,
status board, Gen3/Gen4, LLM inference, performance, P2P, glossary,
troubleshooting and Gen2 pages. Also closes open question 1.7: reproduction of
Gen2 at x16 is settled, so it now asks for the one thing still missing, error
counters after days rather than 90 minutes.
History pages are deliberately untouched. The branch lineage analysis, timeline
and dead-ends record describe what happened on branches, and rewriting them
would falsify the history rather than update the guidance.
2b7e527
Drop the Gen2 x16 rarity note
The table rows already state the measured figures. An alert arguing the
configuration is normal protested the point rather than making it.
4140c53
Stop framing the capacitor mod as rare or risky
The page counted published bandwidth captures and then presented that count as
if it measured how often the mod succeeds. It does not. The rework is routine
0402 work, a Shenzhen supplier sells it as a service, cards ship pre-modded, and
Gen2 is in cmpunlocker master. Gen2 at x16 is what you get when you apply both,
not a rare event. The measurement gap stays documented, described as a gap in
the record rather than as doubt about the configuration.
- Adds a section addressed to a repair shop being asked to quote the work
- Leads the difficulty section with what the job is: 24 0402 caps, labelled
uncramped pads, no BGA, no stencil, no firmware step
- Drops the claim that improper soldering is likely to brick cards, which
contradicts this page's own evidence that partial joints give x8 and are fixed
by reflowing; the documented route to a dead board is over-preheating
- States explicitly that an imperfect result is recoverable
e5c57b4
Add the local-repair-shop route to the cap mod pricing
The Shenzhen 1000 RMB service is not the floor. One builder had cards done
locally for about $50 each. The reported obstacle was not the soldering but
finding a shop willing to take the job at all: more than ten refused before one
agreed to just fit the parts.
175ae1a
Remove all 113 Superseded blocks
These recorded community theories that were later disproven, but the bare
'Superseded' label read as 'this wiki page was superseded' rather than 'this
claim was refuted', and the ambiguity was a bigger liability than the record
was worth.
Also cleans up after the removal: drops five section headings whose entire body
was a Superseded block, renumbers Procedures-Troubleshooting section 9 so it no
longer starts at 9.2, and updates the convention tables on Home and
Appendix-Methodology, which still documented mkdocs '!!!' syntax and a
Superseded convention that no longer exists.
ea3ed71
Gen2 is shipped in master; record the unverified P2P report
PCIe Gen2 merged to cmpunlocker master on 2026-07-29 in commit 2e0a2c02, and
master's README now lists it as working. The wiki still described it as living
only on unreleased branches, across seven places.
Separately, one builder reports the layered aikitoria P2P patch half-works on
GA100: peer DMA at 6.25 GB/s, peer synchronisation broken, so NCCL and vLLM's
custom all-reduce both hang. Single unreproduced rig, so it is recorded as an
explicitly flagged unverified claim with the test that would settle it.
Also reconciles the status legend: drops the now-orphaned 'Working (branch)'
and adds three labels that were in use but never defined.
83067cb
Convert 555 mkdocs admonitions to GitHub alerts
GitHub wikis do not implement mkdocs '!!!' syntax, so every admonition on every
page rendered as literal '!!! warning ...' text with its body flattened. GitHub
does render its own alert blockquotes, verified on this wiki before converting.
mkdocs's eight types collapse onto GitHub's five (note/info/abstract/question ->
NOTE, tip/success -> TIP, warning -> WARNING, danger -> CAUTION). Custom titles
have no slot in a GitHub alert, so each becomes a bold first line inside it.
Also removes the temporary render-test page.
96cd905
temp: test alert rendering
256ea29
Fix field-report image links: use wiki raw URLs
Relative image paths resolve against /wiki/<Page> and 404. The wiki repo's raw
host is the only form GitHub actually serves.
0b99947
Add community field report: four cards at Gen2 x16 on MAASJ105SB7224KFCA01
A community member supplied photographs and logs from four CMP 170HX cards
modded with 24x Taiyo Yuden MAASJ105SB7224KFCA01 (220 nF, 6.3 V, X7R, 0402).
All four train PCIe 2.0 x16 with zero AER errors.
- New field report section with photos, lspci captures, AER counters, bandwidth
- Confirms only the AC-coupling capacitors were needed: stock VBIOS, no VRM
phases, no strap moves, no backup flash chip, no other footprint touched
- Records board revision 180-11001-DAAA-B35, not previously documented
- Gen2 x16 status: Experimental (observed once) -> Working, lightly reproduced
- Documents that lspci may print 5GT/s (overdriven) when the link speed is
forced without rewriting LnkCap; this is normal, not a fault
ab2000c
Correct PCIe AC-coupling capacitor voltage rating: 6.3 V, not 16 V
The BOM said "16 V or higher" with 10 V given as the loosest reported floor.
Both figures were wrong.
Ground truth is page 2 of the circulating a100-unlock.pdf, whose Component
Inventory reads:
24pcs - 0402 Capacitor 6.3v .22uF - MAASJ105SB7224KFCA01
Twenty-four parts, matching the capacitor count exactly, from the x16 lane mod
that demonstrably worked.
Why this was missed: all four pages of that PDF are flat scanned images with no
text layer, so the corpus extraction pulled zero characters from it and the
wiki was built from the chat logs alone.
Where the wrong numbers came from:
- 16 V was an incidental property of the Samsung CL05B224KO5NNNC that a builder
bought as a substitute. It was never a requirement. That part is now listed
as a reported-working substitute rather than the canonical one.
- 10 V traces to the CMP 100 guide, a different card, and was mislabelled here
as a 170HX owner's floor. Removed.
- The only 16 V part in the PDF's inventory is the 10 uF 0805 electrolytic for
the VRM, unrelated to the PCIe lanes.
Also records why 6.3 V is sufficient: these sit in series with a differential
pair and PCIe bounds transmitter DC common mode to 3.6 V.
e851b7b
Publish CMP 170HX wiki: 55 pages plus sidebar
Generated from the MkDocs source in docs/. Page content is byte-identical to
the source apart from link targets, which are rewritten from relative paths
to flat wiki page names.
Navigation is in _Sidebar.md. Start at Home.
Regenerate with to_github_wiki.py rather than editing pages here directly,
so the wiki and the source in docs/ do not drift apart.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2f01d1d