PCIe x16 capacitor mod

PCIe x16 link width (capacitor mod)
Proven 8 GB and 10 GB
Unlock build	hardware mod: 24 x 0402 220 nF capacitors
Steps	PCIe x16 capacitor mod
Source	link
Confirmations	Reproduced by @anon33anon, 2026-10-15 · details

The CMP 170HX trains its PCIe link at x4 because the AC-coupling capacitors on 12 of its 16 lanes were never fitted at the factory. Soldering in the 24 missing capacitors gives a full x16 link. It is the only hardware step in the usual unlock, it needs no software of any kind, and it works the same on the 8 GB and the 10 GB card.

The mod changes link width only, never the PCIe generation. A modded card with no unlock installed reports Gen1 x16; with cmpunlocker it reports Gen2 x16. Generation is a separate, software problem: see Gen2 link retrain and PCIe Gen3 on the 10 GB card.

What the card ships with

The card is wired for x16 but trains x4

All 16 lanes are routed on the PCB; only the AC-coupling capacitors for lanes 4-15 are missing. The lanes are not fused off: OPT_PCIE_LANE_DISABLE (0x00820394) reads 0x00000000 on every card probed. A stock card shows LnkCap: ... Width x16 but LnkSta: Speed 2.5GT/s, Width x4 (downgraded); the "(downgraded)" marker is what a link with no signal on 12 lanes produces.

Reported 8 GB and 10 GB
Source: repo github.com/Consensus-Protocol/cmp170hx by Consensus-Protocol on GitHub, posted by @rrickk420, 2026-10-09

Any x16 link needs the mod; x4 is the most without it

"Any pcie gen with 16x needs cap mod." Without the mod the maximum width is x4, at any generation, including Gen3 on a 10 GB card.

Proven 8 GB and 10 GB
Source: @lxtend, 2026-10-17
Source: @djpandaman, 2026-10-17

Gen3 x16 and Gen4 on the 10 GB card also need the caps

Humvee55's 10 GB cards run Gen3 x16 after the mod. Without the caps, mrpandagod could not get a Gen4 link: it kept falling back to 8.0 GT/s.

Reported 10 GB
Source: @djpandaman, 2026-10-08
Source: @scallion_9883, 2026-10-16

Parts

24 x 0402 220 nF X7R capacitors

You need 2 capacitors per lane for 12 lanes: 24 parts, 0402 size (about 1.0 x 0.5 mm), 0.22 uF (220 nF), X7R. The value comes from the A100 reference schematic (GA100-883, P1001-B02, page 3, PCIe connector). Order at least 30 to cover losses. Rated 16 V in Amogh's write-up; Consensus-Protocol gives 6.3 V or higher.

Confirmed parts:

Samsung Electro-Mechanics CL05B224KO5NNNC (DigiKey 1276-1176-1-ND, 16 V). @globalapathy bought 200 for

 under 5 USD; Mouser and LCSC also stock it.

Taiyo Yuden MAASJ105SB7224KFCA01 (6.3 V).

Reported 8 GB and 10 GB
Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub
Source: @globalapathy, 2026-10-11
Source: @gpmickey, 2026-10-11

No strap resistors, no VBIOS chip

Only the lane capacitors matter for width. No strap resistor is moved and no backup flash chip is added. A circulating "A100 hardware restore" PDF lists a Winbond backup VBIOS chip and other parts; none of them are needed for x16.

Reported 8 GB and 10 GB
Source: repo github.com/Consensus-Protocol/cmp170hx by Consensus-Protocol on GitHub, posted by @rrickk420, 2026-10-09

Where on the board

The C1100-C1350 pads beside the gold fingers

The empty pads carry reference designators in the C1100-C1350 range (for example C1120, C1125, C1130, C1135), two per differential pair. They sit on the PCIe lane routing between the fan-out traces and the edge-connector gold fingers, on the outer surface, uncramped and with visible copper. Cross-check the schematic against your own board to find exactly which pads are empty. The card has to come out of its shroud first: follow the teardown on 170th Street (see Cooling and power).

Reported 8 GB and 10 GB
Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub, posted by @markshowa, 2026-10-18

The leftmost group by the mounting hole is the hardest

With hot air, the leftmost capacitor group next to the mounting hole takes the most heat to do. A solder stencil for the mod is in the pinned messages of the hardware channel, but you need a laser cutter to make it.

Reported 8 GB and 10 GB
Source: @globalapathy, 2026-10-27
Source: @globalapathy, 2026-10-27

Before you start

Test the card stock first. The acceptance tests do not need the mod, and a card you have soldered is harder to return.

 See Acceptance checklist.

Open the card once: repad and repaste at the same time if you plan to.

 Source: @gpmickey, 2026-10-11

If you have never soldered 0402 parts, practise on a dead board first. Several members learned on these cards.

Tools

What members used

A hot-air station and magnification are the core; a fine iron works too. Reported kits:

@neapopa: a hot-air station and a magnifying glass, about 100 USD for both, enough for 8 cards with little prior

 experience.

@civiloid: an ATTEN hot-air station, a Pinecil iron (any 60 W or more iron with SMD tips), an Andonstar microscope

 and tacky flux. He would rather use a microscope with a camera than work off a screen.

@globalapathy: an Atten ST-862D (1000 W) works, but slower than a 1400 W station; Sequre hot tweezers to pull a bad

 capacitor off.

@_arcanum_: microsoldering tweezers are essential, the capacitors are almost impossible to handle otherwise; clean

 with isopropyl alcohol, foam swabs, a brush and wipes.

@gpmickey (twelve cards to do): an upgraded soldering iron, a digital scope and a hot-air station on top of the DMM

 he had, plus extra flux, solder and braid so he would not run short. @globalapathy added a 50 MHz scope meter for
 under 100 USD.

More tools are on Tools.

Reported 8 GB and 10 GB
Source: @globalapathy, 2026-10-11
Source: @gpmickey, 2026-10-11
Source: @gpmickey, 2026-10-11
Source: @civiloid, 2026-10-11
Source: @_arcanum_, 2026-10-11
Source: @globalapathy, 2026-10-27
Source: @neapopa, 2026-10-27

Steps, as reported

Two methods work. Both start from the same advice: flux and a clean tip, and fix one side of every capacitor before doing the other side. Source: @shirogeek, 2026-10-13

Hot air (neapopa, globalapathy)

Preheat the board.
Hold each capacitor with tweezers in contact with its two pads.
Heat with hot air until it bonds, then remove the heat. Use the medium nozzle at low air (globalapathy runs 15 percent)

 so the parts do not blow away.

Nudge the capacitor to confirm it is soldered, then move to the next.

@neapopa added no solder: the solder already on the pads was enough. His first card took 2 hours, the last about 30 minutes.

Reported 8 GB and 10 GB
Source: @neapopa, 2026-10-27
Source: @globalapathy, 2026-10-27

Iron (cmurrayis, Waka324)

@cmurrayis removed the factory solder first, re-applied solder, then fitted each capacitor one at a time with soldering tweezers; about an hour per card. Any wobble shifts the alignment, and it is hard to see whether both ends wetted. Clean the flux off afterwards; @civiloid advises a second round of cleaning, since every flux has long-term side effects. @gpmickey: done well, a soldering iron is a solid alternative to hot air, and there is very little risk on this part of the board.

@cmdrwaka324 (Waka324) found his own iron-and-wick method easier than the video's:

Apply flux.
Use wick and a large flat tip to soak up the original solder (important, or hand work is difficult).
Place a new bead on one side, going up and down the caps, by tinning the tip and dabbing the pads.
With good tweezers, pin one side of each cap with a fine-tipped iron: heat the deposited blob and slide the cap in.
Do the other side, patiently, so solder wicks to both the pad and the cap.
Check by briefly reheating the first side of every cap and watching for movement; if good, no cap should move.
Clean and done.

Reported 8 GB and 10 GB
Source: @gpmickey, 2026-10-13
Source: @cmurrayis, 2026-10-22
Source: @civiloid, 2026-10-20
Source: @cmdrwaka324, 2026-10-05

Remove the factory solder first, or not

Contested

Remove the factory solder with wick first, then re-apply; on the Consensus-Protocol procedure, wick off the lead-free

 factory solder and use leaded solder. Source: @cmurrayis, 2026-10-22

No extra solder needed; the solder on the pads is enough with hot air. Source: @neapopa, 2026-10-27

Video guides

The video "How to Solder Components Without Blowing Them Away (NVIDIA CMP 170HX)" shows the mod done quickly; mrpandagod calls it the best guide, used together with the 170th Street page. Members vouch for the channel's work. https://www.youtube.com/watch?v=HO2T4TgNzlQ

Reported 8 GB and 10 GB
Source: @djpandaman, 2026-10-05
Source: @_orrorin, 2026-10-13

Verify

sudo lspci -s <port> -vvv | grep LnkSta

Read LnkSta, never LnkCap (LnkCap advertises x16 even on a stock card). After a good mod the width field loses "(downgraded)":

State	LnkSta
Stock, no unlock	`Speed 2.5GT/s, Width x4 (downgraded)`
Mod, no unlock	`Speed 2.5GT/s, Width x16`
Unlock, no mod	`Speed 5GT/s, Width x4 (downgraded)`
Mod and unlock	`Speed 5GT/s, Width x16`

Also check the slot is electrically x16, and that the card is fully seated: one card in a 4-card rig trained only x4 or x8 until a spacer pushed it fully into the slot.

Reported 8 GB and 10 GB
Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub
Source: @nestaa51, 2026-10-15
Source: @cibernox, 2026-10-06

Results

x4 to x16, confirmed on many cards

Amogh documented the first public success on a 170HX in April 2026: Speed 2.5GT/s, Width x4 (downgraded) before, Speed 2.5GT/s, Width x16 after. Since then many members report x16, among them @anon33anon (three cards), @neapopa (8 cards) and Humvee55 (10 GB cards at Gen3 x16). See the confirmation below.

Reported 8 GB and 10 GB
Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub
Source: @anon33anon, 2026-10-15

Bandwidth

Stock Gen1 x4 is about 1 GB/s (0.85 GB/s measured). The mod alone gives Gen1 x16, about 4 GB/s. With the Gen2 unlock, Gen2 x16 measured 6.63-6.67 GB/s once. Morrowmake's P2P fork measured about 6.65 GB/s each way between cards at Gen2 x16 (see P2P).

Reported 8 GB and 10 GB
Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub, posted by @markshowa, 2026-10-18
Source: repo github.com/Consensus-Protocol/cmp170hx by Consensus-Protocol on GitHub, posted by @rrickk420, 2026-10-09

Prefill gains a lot, single-card decode barely moves

On a single 40 GB card running Strata, going from x4 to x16 gave about 4x prefill but only about 3 % more decode.

Reported 10 GB
Source: @exile_1ck, 2026-10-07
Source: @wakeupmbrc1, 2026-10-07

Multi-card recipes: the mod is behind the top numbers

@segmond ran Morrowmake's GLM recipe on 4 cards without P2P and without the mod and was "not even close" to the

 published numbers; @cibernox: "the 16x mod is doing a lot for the high speed".

@greencandle9753: to match his DeepSeek V4.1 Flash numbers (150-296 tok/s single user) you have to do the x16 mod.
@zingy90210 (2x 8 GB, no mod) concluded the mod is what makes @nestaa51's 2-card GLM Flash setup fast.
The mod is not required for TP4: Finishtilt14 runs the 4-card GLM recipe without it in TP or PP. TP=2 works well with

 it, and @cibernox calls it "borderline mandatory" for TP=2.

Reported 8 GB and 10 GB
Source: @cibernox, 2026-10-13
Source: @finishtilt14, 2026-10-13
Source: @cibernox, 2026-10-20
Source: @greencandle9753, 2026-10-22
Source: @zingy90210, 2026-10-24

Same link bandwidth as a Gen3 unlock, without a fuse burn

Gen2 x16 carries what Gen3 x8 would, so for tensor parallel the mod gives the gain a Gen3 unlock would, with a soldering iron instead of an irreversible RIR fuse burn. Ivan's estimate for TP4 single user, 280-355 tok/s toward 440-480 tok/s, is a projection, not a measurement.

Theory 8 GB and 10 GB
Source: @darkmaniac7, 2026-10-19 (lab notes)

No benefit for mining

One member modded one card and skipped the rest because the mod gives no mining advantage.

Reported 8 GB and 10 GB
Source: @port00, 2026-10-22

When it does not come up x16

x8 means a bad joint, not a hardware limit

Width falls back to the next legal value instead of failing: 12 of 24 good capacitors gives x8, and so does any cold, missing or bridged joint. A card that trains x8 or x4 after the mod still works at that width while you fix it.

@bigfoot1144: still x4 after the first soldering, x8 after cleaning the caps.
@anon33anon: third card stuck at x8 although continuity from each cap top to the slot looked fine and nothing was

 shorted; after redoing the suspect caps it came up x16.

@cmdrwaka324: on his early cards a bad cap job showed as x16 dropping to x8, until he refined his process.

Reported 8 GB and 10 GB
Source: @bigfoot1144, 2026-10-13
Source: @anon33anon, 2026-10-15
Source: @cmdrwaka324, 2026-10-16

How to find the bad capacitor

Test continuity between each capacitor and the gold finger it should reach, not just from the cap top to the slot.

 @globalapathy notes the top pad cannot be probed without scraping the trace, so a cap that measures fine from the top
 can still be lifting.

With a DMM, check for shorts between the signal lines.
Under the microscope, tilt the card to see under the ends of each capacitor; dull joints are cold joints. Reheat them

 with good SMD flux.

Or tap both ends of every capacitor with the iron; a cap that is not making contact shows up quickly.

More link problems after the mod: Card not detected or BAR allocation fails.

Reported 8 GB and 10 GB
Source: @globalapathy, 2026-10-13
Source: @ghostinthedatacenter_87281, 2026-10-15
Source: @chonsworth, 2026-10-15
Source: @cmdrwaka324, 2026-10-15
Source: @globalapathy, 2026-10-27

Full x16 can expose a power problem

Five cards on an external GPU switch baseboard, cap modded, dropped off PCIe with lane and IOMMU errors on large prompts; the logs showed the first card drawing 260 W. Lowering the power limit from 250 W to 180 W fixed it (GLM 5.3 Flash at 380k context). See Cooling and power.

Reported 8 GB and 10 GB
Source: @moliminous, 2026-10-19

Risks

Hot air blows the parts away

Warning At normal air flow a hot-air nozzle blows 0402 capacitors (yours and the board's) off the pads; the board components do not stick until it has warmed up. Use the medium nozzle at low air (about 15 percent), preheat, and keep hot tweezers or an iron ready to pull a misplaced part. @_orrorin shields neighbouring parts with aluminium tape.

Source: @globalapathy, 2026-10-27
Source: @_orrorin, 2026-10-13

No card is reported killed by the mod

Across the Discord history (2026-10-04 to 10-28) nobody reports a card damaged by the mod itself. A failed attempt that does not damage other components is easy for a technician to fix, and the card keeps working at x4 meanwhile.

Reported 8 GB and 10 GB
Source: @_orrorin, 2026-10-13
Source: @gpmickey, 2026-10-13

A card from a repair technician arrived dead

@exile_1ck had his card's x16 mod done by a local GPU repair technician in under an hour; he found him on Facebook Marketplace as a buyer of broken GPUs, which gave him a repair history to check. The technician takes no online orders and eXile keeps his name private, so he is not listed here. Later eXile bought a non-working 10 GB card from the same technician (taken in from a customer, history unknown); it does not finish booting and appears blocked by GPC6 (see Recovery). Nothing links its failure to the cap mod.

Reported 10 GB
Source: @exile_1ck, 2026-10-04
Source: @exile_1ck, 2026-10-11

Having it done

Repair shops and prices

Shops that do board-level GPU work treat this as basic: fitting capacitors, far less work than a reball. Ways members found one: search for "microsolder" on a local review site and call around; or look for people buying broken GPUs to repair and resell, who have review histories. Quotes reported (October 2026), all for a seller doing only the soldering:

45 USD per card, customer supplying the 0402 caps (California).
120 CAD per card (British Columbia); 160 CAD per card (Edmonton).
65 EUR per card for 13 cards (EU); 300 AUD for one card (Australia); about 900 USD for 8 cards.
175 USD per card with the customer supplying the caps; the shop would not guarantee the GPU would be "repaired".
In China about 50 CNY (under 10 USD).
A paid service at about 500 USD was judged too expensive when the gear costs under 100 USD.

Ailfond ships cards already modded, about 2,600 USD DDP (October 2026); members asked whether anyone does the mod cheaper than Ailfond, with no clear answer.

Reported
Source: @kellertreppe123, 2026-10-04
Source: @nikhilr1696, 2026-10-04
Source: @cibernox, 2026-10-19
Source: @68x0, 2026-10-06
Source: @gpmickey, 2026-10-06
Source: @exile_1ck, 2026-10-04
Source: @gnasherrr, 2026-10-21
Source: @lxtend, 2026-10-21
Source: @cmurrayis, 2026-10-22
Source: @reala7096, 2026-10-22
Source: @impactic, 2026-10-22
Source: @danske1, 2026-10-27
Source: @endsjeremy, 2026-10-16

Community members who offered the service

Members who said they would do the mod for others (October 2026):

@gmzombie (GMZombie007), US, 20+ years electronics work: offered it through an eBay listing so buyers get a receipt,

 about 50 USD plus shipping, discount for multiples. Source: @gmzombie, 2026-10-04

@tokukarin (Tokukarin): offers the service and did all his own cards. Source: @tokukarin, 2026-10-04
@neapopa (Nea Popa): would do it for 50 USD with caps supplied; did 8 cards himself. Source: @neapopa, 2026-10-27
@ikyo (Ikyo) knows a US contact last quoted at 100 USD per card. Source: @ikyo, 2026-10-04
GPU Lab (gpulab.net, product id 16) lists the mod as a paid service at about 500 USD; @endsjeremy argued that is too

 much when the gear costs under 100 USD. Source: @globalapathy, 2026-10-16

 Source: @endsjeremy, 2026-10-16

Shops say no

Owners report being turned away by many shops before finding one that would simply do the work; the job is 24 capacitors on labelled pads, with no diagnosis and no firmware step. Hand the shop the parts and that description. Local options are thin in some countries: an EU buyer weighed a 1,900 EUR unmodded card partly because nobody local could do the mod.

Reported
Source: repo github.com/Consensus-Protocol/cmp170hx by Consensus-Protocol on GitHub, posted by @rrickk420, 2026-10-09
Source: @shudza, 2026-10-19

Status

x4 to x16 is reproduced by many owners on both cards, so the unlock shows Proven through the confirmation below.