Jump to content
Toggle menu
Toggle preferences menu
Toggle personal menu
Not logged in
Your IP address will be publicly visible if you make any edits.

Unlock without bricking

From CMP Wiki

Warning Read this before you burn anything. A CMP 170HX floorsweep is per die. Another card's fuse words are not your plan. OTP bits burn from 0 to 1 and do not clear. Most of the glitcher tooling in circulation in November 2026 is a few days old, much of it generated by a language model, and tested on one card. Two models agreeing is not a check against your own dump.

This page is the order of operations: what is reversible, what is permanent, and the mistakes that have already stuck cards. The feature pages (Full SM unlock, RIR fuse burn, PCIe Gen3 on the 10 GB card, PCIe Gen4 on the 10 GB card, Memory unlock on the 10 GB card, NVLink, ECC) keep the history. The 10 GB how-to with registers, record lists, readback and the burn steps is SW-IFF unlock on the 10 GB card. Start at Which card do I have.

Which path your card is on

  • 8 GB card (PCI ID 10de:20c2). The redundancy rows were burned at the factory. cmpunlocker is the software path: 64 GB and 74 SM. There is no second software step from 74 to 78. The "+4 SM" after a hardware unlock is the same RECONFIG override (see below). Anything past that is the voltage-glitch route, then permanent FF records. That route is Reported, and the lab has not reproduced it.
  • 10 GB card (PCI ID 10de:2082). It has empty RIR rows. One RIR burn opens the software-override gate. No glitcher is required for that burn, and on the two cards below no glitcher was required to open the FF write window afterwards. Prove the mask in volatile software on bare metal, then burn only that mask if you need it to survive a reboot or a passthrough reset.

Reported 8 GB and 10 GB
Source: @killred, 2026-11-16
Source: @darkmaniac7, 2026-11-17 (lab notes)

Two fuse systems

RIR and FF are different things in BAR0, not two names for one burn.

  • RIR (redundancy rows). On the 10 GB card the public tool writes only rows 0 to 7, through the RIR aperture (RWL=1). That is the gate burn. It is the wrong aperture for a floorsweep. Pointing a main-array write at RWL=1 is a known way to brick the card.
  • FF (fuseless-fuse records). Append-only. A later record supersedes an earlier one, the earlier row stays spent, and one row can hold several fields so neighbouring bits have to be preserved. The die records defective and disabled separately. Clearing a disable bit does not repair a defective unit. First posted on the clean room server on 2026-10-03 by @akuchen_66498, in a plan he put up for review before touching OTP: "Last-writer replaces the whole 16-bit block, so the appended block has to reproduce every other bit correctly."

Reported 8 GB and 10 GB
Source: @akuchen_66498, 2026-10-03 (CMP 170hx Clean Room Discussion server)
Source: @killred, 2026-11-16
Source: @galygious, 2026-11-16

The glitch, when you need one, opens the software override so the FF window can be read and written. It does not itself change a permanent bit.

Reported
Source: @galygious, 2026-11-16

Where the FF rows start and end

Kate's blog, as @killred quoted it, puts regular fuses in rows 0-221 and the fuseless-fuse (FF) records in rows 222-487, with new records going after row 222 into an empty row; his own dump matched the blog for the first rows. A word with bit 5 set marks the end of the FF rows and the start of the IFF region ("bit5=0 FF, bit5=1 IFF", @stsoftware_88487), and @bigptoughneigh remembers that the FF start and the IFF start may differ per card. On one 8 GB card @port00 measured what the boot parser read: rows 222-302 were consumed, rows 303-487 were not, and records later burned into regular rows 128-131 were ignored. An AI-generated summary that @68x0 relayed from @petrikrohn_26497 puts the cell-backed rows at 32-221 and ends the FF pool at the IFF start or at 8 consecutive zero rows (AI-generated analysis, not measured). Where the parser stops has not been settled by a measurement, so read your own card before you pick a row. Row counts and the decode: Fuses and floorsweep maps#FF records and the IFF window: what members measured on 8 GB cards.

Reported 8 GB and 10 GB
Source: @stsoftware_88487, 2026-11-10 (CMP 170hx Clean Room Discussion server)
Source: @bigptoughneigh, 2026-11-10 (CMP 170hx Clean Room Discussion server)
Source: @killred, 2026-11-16
Source: @port00, 2026-11-16
Source: @killred, 2026-11-17
Source: @68x0, 2026-11-17

SBFE, not DSOV alone, opens the FF rows

@matstatman asked on 2026-11-16 whether DSOV=0 is needed at all, or only to get SBFE=0. @mellojello_ answered that DSOV=0 is not needed in the traditional sense, since fuses can be burned without it, but "SBFE=0 is required as otherwise you cannot burn any rows within the FF space". Getting one without the other does not seem possible, and glitching straight at SBFE had not been tried. What gates a main-array burn is still argued: Fuses and floorsweep maps#What gates a main-array burn (contested).

Reported 8 GB and 10 GB
Source: @matstatman, 2026-11-16 (CMP 170hx Clean Room Discussion server)
Source: @mellojello_, 2026-11-16 (CMP 170hx Clean Room Discussion server)

Without the gate only the field-programmable rows take a burn

On stock-gated cards members could burn only the field-programmable fuse (FPF) rows, about 191 to 221. @n2dcrypt burned a test value into one row of that range on an 8 GB card to prove the machinery works, but "everything outside of that window was no land": a burn at row 302 was accepted and completed without error, and nothing was programmed. @stsoftware_88487's attempts at FF rows 293 and 351 failed the same way. A burn that lands in the FPF rows proves the write path, not an unlock, and an accepted command is not a programmed row: read every row back. Details: Fuses and floorsweep maps#Which rows take a burn: the field-programmable rows, and a silent wall in the FF rows.

Reported 8 GB and 10 GB
Source: @stsoftware_88487, 2026-10-07 (CMP 170hx Clean Room Discussion server)
Source: @n2dcrypt, 2026-11-03 (CMP 170hx Clean Room Discussion server)
Source: @stsoftware_88487, 2026-11-05 (CMP 170hx Clean Room Discussion server)
Source: @n2dcrypt, 2026-11-06 (CMP 170hx Clean Room Discussion server)
Source: @n2dcrypt, 2026-11-07 (CMP 170hx Clean Room Discussion server)

Safe order on a 10 GB card

  1. Install cmpunlocker and confirm the 40 GB profile and 74 SM. Do not burn yet.
  2. Read the defect fuses and the disable fuses. Enable a unit only when it is disabled and not defective. The SM count you are allowed to aim at is 2 * (8 - bad TPCs) summed over GPCs you are allowed to turn on. A GPC whose TPCs are all defective stays off. 128 SM is a die with no bad TPC. 126 is one bad TPC. Do not turn a bad TPC on to make the number bigger.
  3. Burn the RIR gate with 170hx-rir-tool (or the same record that tool writes) into a blank RIR row. The tool refuses a row that is not blank. This is the one OTP step that does not need a glitch on this card. An FLR does not re-sense it: DSOV (0x820084) can stay 1 after echo 1 to the PCI reset node. A secondary bus reset, or a full power cycle, is what makes DSOV read 0. A driver bind then unbind opens the fuse PLMs (0x8200e0 and 0x8200f4). They stay open across the unbind.
  4. Prove the floorsweep with SW-IFF on bare metal, device unbound, then an FLR, then bind the driver. Zero Xid, the CUDA SM count you calculated, defective TPC status bits still set, memory size unchanged. Power cycle returns the card to stock. That is the point of doing it this way.
  5. Only after that proof, and only if you need the mask to survive a reboot or a VM reset, append FF records of that same mask into blank rows inside the parsed window. Write the TPC and PES records first and the GPC record last. Read every row back. Stop on a mismatch. Do not retry a failed write. CRC (0x820070) stays 0x55.
  6. Leave memory, PCIe generation, NVLink, the device id and the VBIOS out of that burn. On the cards below, none of those is a pure fuse flip (see the last section).

Reported 10 GB
Source: @darkmaniac7, 2026-11-17 (lab notes)

Members gave the same order on the clean room server from early October. @petrikrohn_26497 (2026-10-03): on a 10 GB card the right way is to enable the software override with one available RIR record and use the software method; "When you are absolutely 100% sure what you want to change, use the spare fuse rows to rewrite the FF fuses." He repeated on 2026-10-17 to use the software override first, with no need to burn FF fuses, and on 2026-10-20 that burning the two RIR rows is the risky part. @stsoftware_88487: one RIR record gives a 10 GB owner SW-IFF and control of all fuses. Asked whether 170hx-rir-tool is safe, @djpandaman said to take a second to work out what you burn, why, and how it was calculated, since everything else is reset by a cold boot; @68x0 told a newcomer to be careful with it, and @ghostinthedatacenter_87281 did only the one burn, for the software override.

Reported 10 GB
Source: @petrikrohn_26497, 2026-10-03 (CMP 170hx Clean Room Discussion server)
Source: @petrikrohn_26497, 2026-10-03 (CMP 170hx Clean Room Discussion server)
Source: @stsoftware_88487, 2026-10-03 (CMP 170hx Clean Room Discussion server)
Source: @innospark, 2026-10-16 (CMP 170hx Clean Room Discussion server)
Source: @djpandaman, 2026-10-17 (CMP 170hx Clean Room Discussion server)
Source: @petrikrohn_26497, 2026-10-17 (CMP 170hx Clean Room Discussion server)
Source: @ghostinthedatacenter_87281, 2026-10-17 (CMP 170hx Clean Room Discussion server)
Source: @petrikrohn_26497, 2026-10-20 (CMP 170hx Clean Room Discussion server)
Source: @68x0, 2026-10-23 (CMP 170hx Clean Room Discussion server)

Warning 8 GB and 10 GB Where a permanent record goes. @samuel_hayden. (2026-11-17), right after warning "If you yolo ffcalcs you will brick your card": new FF records "have to be burned in the empty space", and only "with a buffer above them so you cannot interfere with any of the boot ffs"; "you should leave a clear empty space at the end for reverts"; and leave room to re-disable SMs that turn out to be bad. Count your free rows first (see #Do not do these).

Source: @samuel_hayden., 2026-11-17
Source: @samuel_hayden., 2026-11-17
Source: @samuel_hayden., 2026-11-17
Source: @samuel_hayden., 2026-11-17

What SW-IFF actually does

The overlay lives in volatile SEC2 SRAM, pointed at by IFF_RECORD (0x820050). It has to replay at fuse sense, before devinit. A late OPT write, after the driver is up, only fixes TPCs inside GPCs that already came up, which is why people get stuck near 80 SM and then see FECS / HOST-PRI timeouts when they poke a whole GPC. SW-IFF is how the whole GPC is on in time for init.

On bare metal, an FLR keeps that SRAM, so the records replay. On VMware ESXi PCI passthrough, both flr and link reset methods wipe it (IFF_RECORD returns to the stock pointer, SRAM reads as zeroes) before replay. The same arm-and-FLR that fails in that guest works on bare metal. Other hypervisors were not tested. A permanent FF burn is the expected route for passthrough, because the fuses are sensed again from OTP on every reset, but a burned card has not yet been tested in a guest. Volatile SW-IFF does not survive the passthrough reset.

While EN_SW_OVERRIDE (0x820040) is 1, a fuse-controller read of a row inside the IFF window returns the SRAM slot, not the fuse array. Unwritten SRAM looks like noise and will not match the stock window. Read the window with 0x820040 at 0. If a read returns 0xBADFxxxx, stop and cold boot.

On two 10 GB cards the stock window was the same: IFF_RECORD = 0x01fb01e8, rows 488 through 507 byte-identical, and the same fingerprint the tool expects. The floorsweeps differed (GPC disable 0x64 against 0x89), so on these two cards the stock stream is not keyed per die. 0x000D56BD is the first stock record, not the override record. The applier prepends EN_SW_OVERRIDE=1 itself. WRITE_DW (op 0x3D) is followed by one data word. MODIFY_DW (op 0x3C) is followed by two (mask, then value). That is 9 op words in 20 rows.

Reported 10 GB
Source: @darkmaniac7, 2026-11-17 (lab notes)

Two cards, defect-preserving, then permanent

Both cards are Samsung HBM, PCI ID 10de:2082, driver 615.71.09, 10 GB profile. Baseline 74 SM and 40 GiB.

  • Card A. Gate already burned. SW-IFF to the defect-preserving mask, FLR, driver bind, no Xid. CUDA reported 122 SMs (5 of 5 sustained checks). A 38 GiB memory test, 92 rounds, 7508 GB moved, 0 errors. Three factory-defective TPCs stayed off. 40 GiB unchanged. A power cycle returned it to stock.
  • Card B. Gate was unburned (DSOV=1). GPC7 is wholly defective, so it stays off. One bad TPC in GPC0 and one in GPC3. Defect-preserving maximum is 108 SM. RIR gate burn, bus reset (not FLR) to make DSOV=0, one driver bind/unbind to open the PLMs, then the same SW-IFF path. CUDA reported 108 SMs. Same memory test, 0 errors.

The FF window preconditions the public writer checks (DSOV=0, CRC 0x55, software override on, secure-burn enables clear, EFP=1, both fuse PLMs open, controller idle) were all reachable in software on these cards. Nothing was programmed in that check. An FLR undoes it.

Both masks were then burned as compute-only FF records (seven records each, TPC and PES before the GPC record, every row read back, CRC stayed 0x55). Memory, device id, PCIe and NVLink records were not written. A cold boot from a settled driver (bound, window closed) came up with the software override off and the counts still there: 122 and 108, 40 GiB, no boot fault. The first attempt to reboot straight out of the open burn window (driver unbound, override on) hung at POST and needed a hard reset. Rebind and let the card settle before you power cycle.

The row words are not on this page. They are this die's encoding of this die's defect mask. Burning them onto another card is how you enable dead units.

Reported 10 GB
Source: @darkmaniac7, 2026-11-17 (lab notes)

40 GiB against 48 GiB

Full GA100 is 12 FBPs at 4 GiB, 48 GiB. A 10 GB card grows only if a disabled FBP is not defective. Compare OPT_FBP_DISABLE (0x820364) with fbp_defective (0x8205cc). Recoverable bits are disable AND NOT defective. If that is 0, every disabled FBP is dead and a memory burn does nothing. Check FBPA (0x820368 / 0x8205d0), FBIO (0x82036c / 0x8205d4) and ROP/L2 (0x8202c4 / 0x8205e8) the same way.

On these two cards the disable masks equalled the defect masks (FBP7+FBP8 on the 122 SM card, FBP9+FBP10 on the 108 SM card). A reversible SW-IFF that forced all 12 FBP disable masks to 0 still enumerated 40960 MiB and threw no HBM training error. The 10 GB driver geometry programs 40 GiB, and the extra sites are dead. Do not put FBP records in an SM burn. 80 GB on a Samsung 10 GB card is a different problem: members describe it as locked by Samsung, not by an NVIDIA fuse.

Reported 10 GB
Source: @darkmaniac7, 2026-11-17 (lab notes)
Source: @68x0, 2026-11-16

Members report 48 GB on some 10 GB cards after the RIR unlock, and only where the sixth stack's FBPs are disabled, not defective. @petrikrohn_26497 (2026-10-20): "The 10GB cards can be unlocked to 48GB, assuming all six HBM2 stacks are not *defective*, but just *disabled*." @_orrorin burned the gate on his cards and reported on 2026-10-21: "my RIR unlocked 10GB definitely, for sure, only has 48GB onboard"; on 2026-11-17 he relayed that Kate unlocked a full card and said 48 GB is the maximum. @mbelli has unlocked two 10 GB cards to 48 GB (2 of his 15, against an average of about 10 percent he had heard) and thinks more would need a Samsung-specific initialisation, possibly over IEEE 1500. @analyticeth: 4 of his 5 10 GB cards went to 48 GB and 128 SMs, and pushing past 48 GB "just aliased back onto another stack already in use". Which cards qualify differs: one of @toffles3179's 10 GB cards has FBPs 6 and 11 disabled but not defective, the other has them disabled and defective, and @mbelli saw 10 GB cards with defects on two FBPs and one with none. In July @bigptoughneigh had read 10 GB cards as having only defective FBPs, which @toffles3179's first card contradicts. @c133 hears that the voltage glitch makes no difference to this on a 10 GB card.

Proven 10 GB
Source: @bigptoughneigh, 2026-07-19 (CMP 170hx Clean Room Discussion server)
Source: @mbelli, 2026-10-04 (CMP 170hx Clean Room Discussion server)
Source: @toffles3179, 2026-10-08 (CMP 170hx Clean Room Discussion server)
Source: @petrikrohn_26497, 2026-10-20 (CMP 170hx Clean Room Discussion server)
Source: @_orrorin, 2026-10-21 (CMP 170hx Clean Room Discussion server)
Source: @c133, 2026-11-15 (CMP 170hx Clean Room Discussion server)
Source: @_orrorin, 2026-11-17 (CMP 170hx Clean Room Discussion server)
Source: @mbelli, 2026-11-17 (CMP 170hx Clean Room Discussion server)
Source: @analyticeth, 2026-11-17
Source: @mbelli, 2026-11-17

@scallion_9883 (2026-11-17): "if the 10gb stacks are 16gb then they're fused on the actual HBM itself, not on the ga100 silicon". If so, no FF burn on the GA100 adds that memory. Not measured.

Theory 10 GB
Source: @scallion_9883, 2026-11-17

Resets

You want Reset that does it Reset that does not
Make a fresh RIR gate visible (DSOV 1 to 0) Secondary bus reset, or a power cycle FLR
Replay an SW-IFF overlay on bare metal FLR. SRAM survives A power cycle. SRAM dies, fuses remain
Replay an SW-IFF overlay inside VMware ESXi passthrough Not observed. flr and link both wiped the SRAM
Keep a burned FF floorsweep Cold boot (proven). ESXi passthrough reset: expected, not yet tested Nothing. It is OTP
Return a volatile experiment to stock Power cycle

Reported 10 GB
Source: @darkmaniac7, 2026-11-17 (lab notes)

On one multi-card host the patched 615.71.09 driver sets NO_BUS_RESET. Unbinding nvidia and rebinding then hangs, and only a cold power cycle brings the cards back. Do not unbind the driver on that host to take a fuse dump. A read-only BAR0 map with the driver bound is the floorsweep read that does not reset the bus.

Warning 8 GB
Source: @darkmaniac7, 2026-11-17 (lab notes)

During glitch runs the card can drop off the bus. Two members, on a Z490 board and an Asus WRX90E, saw config space read 0xffffffff after a glitch: PEX_RST pulses from the Pico, but the root port does not retrain, and only a full power cycle brought the card back; on the WRX90E, BAR resize did not work either. @samuel_hayden.'s similar problem went away after he disconnected PEX_RST (from scope probing). @fallengod9962's whole host locked up during glitching and needed a hard reboot. @ichundes posted a sequence that gets fresh BARs without a reboot: hold PEX_RST low from the Pico, remove the device through sysfs, release PEX_RST, wait about 6 seconds, rescan the bus, then re-enable memory and bus mastering with setpci. More: Voltage-glitch unlock (GA100VFI)#When the link drops: PEX_RST does not retrain the root port.

Reported 8 GB
Source: @samuel_hayden., 2026-11-16
Source: @dr.bubotnic, 2026-11-17
Source: @dr.bubotnic, 2026-11-17
Source: @ghostinthedatacenter_87281, 2026-11-17
Source: @fallengod9962, 2026-11-18
Source: @dr.bubotnic, 2026-11-18
Source: @ichundes, 2026-11-18

The 8 GB card, and the glitcher

cmpunlocker's +4 SM step writes the RECONFIG override (STATUS = DISABLE | DEFECTIVE | CTRL | (RECONFIG & ~OVERRIDE)). It does not clear a defective TPC or a whole disabled GPC. After a hardware FF unlock, the same override is what takes a card from 124 SM to 128 SM when two TPCs were only on RECONFIG. Patrik reported that split on 2026-11-15 (UTC). It is not a new 74-to-78 unlock for a stock 8 GB card.

On 2026-11-17, eight 8 GB cards in one chassis, driver 615.71.09, already showed 74 SMs in CUDA on every card, 65536 MiB, and every RECONFIG-only TPC already exposed. Three whole GPCs and three more TPCs stayed off because they are disabled or defective, not because the override was missing. No further software SM was left.

Reported 8 GB
Source: @._.patrik, 2026-11-15
Source: @zagitta, 2026-11-15
Source: @darkmaniac7, 2026-11-17 (lab notes)

Past that, the reported 8 GB route is Kate Land's voltage glitch (kate-land/ga100vfi), then FF programming. The lab has not built the rig and has not reproduced the 80 GiB result. Treat the following as other people's reports.

What members posted about the method, its tools and its outcomes is collected on Voltage-glitch unlock (GA100VFI); this page keeps what prevents damage.

  • The glitch opens the gate. It is not the floorsweep. You still have to read your own fuses and write your own records. Source: @galygious, 2026-11-16
  • The path people are using is a Raspberry Pi Pico 2 (RP2350) on the board test points (I2C and PEX_RST, not only the pads near U0808/U0809), plus ground. An RP2040 fork exists. Kate's original rig was reported at 1 to 48 hours per card. The Pico path is reported at about 5 attempts per second and about 18 percent success per attempt. One member's longest of about ten tries was about 25 seconds. Source: @c133, 2026-11-16
    Source: @c133, 2026-11-16
    Source: @kingsizexxxl, 2026-11-16
    Source: @kingsizexxxl, 2026-11-16
    Source: @kingsizexxxl, 2026-11-16
  • Solderless pogo fixtures are being built. They are not the default yet. Source: @realsatoshigold, 2026-11-16
  • Confirm the glitch with DSOV, CRC 0x55, and EFP back at 1 after re-sense. One Teensy 4.0 adaptation saw DSOV=0 and CRC 0x55, then EFP stayed 0 and CRC went 0x00 after re-sense. That is not a confirmed glitch. Source: @nestaa51, 2026-11-16
  • Unload nvidia and nouveau before an FF read. The dump does not work with the NVIDIA driver loaded. Source: @killred, 2026-11-16 @68x0 confirmed it: "Yeah, ran into read errors when drivers were loaded". Source: @68x0, 2026-11-16 One read-only script, ga100_fuse_dump.py, is reported to read with the driver loaded; see Fuses and floorsweep maps#Taking a raw fuse dump safely.
  • One Hynix 8 GB card on an RP2350A, stock 150 MHz, was reported open with v_dip=360, gap=150, pulses=3, pw=198, delay=117. That is one rig's report, not a recipe for every board. The same post is the mask-zero brick below. Source: @port00, 2026-11-16

Reported 8 GB

@._.patrik's one-card archive (docs, scripts, a glitcher snapshot, driver patches) is on CMPGlitch fuse research archive, with the author's disclaimer: LLM-generated, one card, fuse burns irreversible, no warranty. The hosted zip is a modified copy. It omits two unlicensed subtrees and four vendor source files. It does not include his RP2040 glitcher firmware or an A100 VBIOS. Do not run it against a card you have not read yourself.

Reported 8 GB
Source: repo github.com/kate-land/ga100vfi by kate-land on GitHub, posted by @._.patrik, 2026-11-16

Before you start

Cautions from the clean room server: do not try voltage glitching when tired, learn the basics of voltage fault injection first, and "If you see magic smoke then you know what not to do" (@cute_sc2, 2026-10-03). A chip can quite easily short itself through a glitch, for example by routing power to a faulty SM, so guard it from destroying itself (@m3gagluk, 2026-10-26). Go in baby steps and first verify that the glitch works, does not damage cards and is reversible (@ins_mniac, 2026-11-11). First identify where the power regulator is, and when and how to glitch (@truvv0r, 2026-11-14). @dalzyyy noted that fuse sensing fails below about 400 mV.

Reported 8 GB
Source: @cute_sc2, 2026-10-03 (CMP 170hx Clean Room Discussion server)
Source: @cute_sc2, 2026-10-03 (CMP 170hx Clean Room Discussion server)
Source: @m3gagluk, 2026-10-26 (CMP 170hx Clean Room Discussion server)
Source: @dalzyyy, 2026-11-11 (CMP 170hx Clean Room Discussion server)
Source: @ins_mniac, 2026-11-11 (CMP 170hx Clean Room Discussion server)
Source: @truvv0r, 2026-11-14 (CMP 170hx Clean Room Discussion server)

Keep the NVIDIA driver off the card

A driver still bound to the card spoils both the glitch run and the fuse read (see the dump note above). On 2026-11-18 @fallengod9962's crash log showed the driver still loaded: xrdp, a remote-desktop server, was holding the GPU, and a reboot through the BMC did not clear it. Fixes members posted: disable the desktop environment (@dalzyyy); "setup an extra boot entry in your bootloader that loads into plain terminal mode without any nvidia driver" (@zagitta); blacklist the NVIDIA driver and update the initramfs so it never loads (@ichundes).

Reported 8 GB and 10 GB
Source: @scallion_9883, 2026-11-18
Source: @fallengod9962, 2026-11-18
Source: @fallengod9962, 2026-11-18
Source: @fallengod9962, 2026-11-18
Source: @dalzyyy, 2026-11-18
Source: @zagitta, 2026-11-18
Source: @ichundes, 2026-11-18

Flash the controller before you connect PEX_RST

Warning 8 GB @ichundes, who glitched two 8 GB cards with a Pico 2: "make sure you flash the glitch firmware first before you connect PEX_RST or it could damage the hardware."

Source: @ichundes, 2026-11-17

Wiring, probes and solder

Members wire four points, all on the back of the card: SDA, SCL, PEX_RST and GND. @ichundes used no NVVDD connection, @ghostinthedatacenter_87281 did not wire it either, and @lia_oc and @zagitta say it is not needed; read the README of dalzyu/pico-ga100vfi closely (its first wiring pictures confused people, and its author took a clearer photo from @ichundes). Nothing has to be removed from the board for this hookup; Kate's original rig removed a resistor. Ground is needed too (@gnasherrr, @wakeupmbrc1), and wires soldered on have to come off again afterwards. Because every point is on the back, the heatsink or water block can stay on; with the stock air cooler, take the shell off and leave the copper heatsink on (@lia_oc).

Solder or probes: @ins_mniac puts four joints at about 2 minutes per card with a good iron, and warns that a pressure jig is risky, because even poor solder joints make the controller link unstable and too much pressure can damage the PCB; @con3po7016 chose soldered wires as the least error-prone. Others use probe holders (PCBite or SensePeek kits) or are building pogo-pin fixtures (@jbranco91, @cmdrwaka324, @ju5t5omedude), and @kingsizexxxl prefers pogo pins to clips. @lia_oc chose probes because "some people have already damaged their cards w/ soldering attempts". The controller does not have to stay on the card, so one Pico can serve several cards in turn (@ju5t5omedude). Kate's PCB is cheap ("a 5 dollar board", @djpandaman); @m3gagluk noted it needs a comparator for a specific voltage, and @mellojello_ answered that there are other routes. Members now use a Pico in its place.

Reported 8 GB
Source: @m3gagluk, 2026-11-11 (CMP 170hx Clean Room Discussion server)
Source: @mellojello_, 2026-11-11 (CMP 170hx Clean Room Discussion server)
Source: @djpandaman, 2026-11-11 (CMP 170hx Clean Room Discussion server)
Source: @ins_mniac, 2026-11-13 (CMP 170hx Clean Room Discussion server)
Source: @ins_mniac, 2026-11-13 (CMP 170hx Clean Room Discussion server)
Source: @ju5t5omedude, 2026-11-14 (CMP 170hx Clean Room Discussion server)
Source: @cmdrwaka324, 2026-11-14 (CMP 170hx Clean Room Discussion server)
Source: @con3po7016, 2026-11-14 (CMP 170hx Clean Room Discussion server)
Source: @ju5t5omedude, 2026-11-14 (CMP 170hx Clean Room Discussion server)
Source: @jbranco91, 2026-11-16
Source: @zagitta, 2026-11-16
Source: @zagitta, 2026-11-16
Source: @kingsizexxxl, 2026-11-16
Source: @wakeupmbrc1, 2026-11-16
Source: @wakeupmbrc1, 2026-11-16
Source: @gnasherrr, 2026-11-16
Source: @lia_oc, 2026-11-16
Source: @lia_oc, 2026-11-16
Source: @gnasherrr, 2026-11-16
Source: @lia_oc, 2026-11-16
Source: @lia_oc, 2026-11-16
Source: @lia_oc, 2026-11-16
Source: @ichundes, 2026-11-17
Source: @ghostinthedatacenter_87281, 2026-11-17
Source: @dalzyyy, 2026-11-17
Source: @dalzyyy, 2026-11-17

Is the ground wire needed?

In the order posted (2026-11-18):

  • @scallion_9883: no separate ground is needed if the Pico is plugged into the same workstation as the GPU, which completes the loop. Source: @scallion_9883, 2026-11-18
    Source: @scallion_9883, 2026-11-18
  • @kingsizexxxl: "Feel like that's an unnecessary risk"; "More ground more better". Source: @kingsizexxxl, 2026-11-18
    Source: @kingsizexxxl, 2026-11-18
  • @fallengod9962, who had forgotten to solder it: his rig "was very flacky for me until i did that getto wiring job". Source: @fallengod9962, 2026-11-18
    Source: @fallengod9962, 2026-11-18

Contested 8 GB

Controller boards

A Raspberry Pi Pico 2 (RP2350) is enough; no Teensy or comparator is needed (@dalzyyy, 2026-11-11 and 11-13), and @scallion_9883 thinks even a Pico 1 is fast enough. The repo points buyers to a Pico 2 or Pico 2 W (@con3po7016). @c133's fork adds RP2040 support (clee/pico-ga100vfi, branch port/rp2040-kb2040), and JosiahWhite/rp2040-glitch-tool is the second controller repo @dalzyyy points to. Other microcontrollers need a port: any MCU with a hardware I2C peripheral and a free GPIO input should do, and an ESP32 needs the code rewritten for its HAL (@samuel_hayden.); a USB I2C adapter such as a CH341A alone is not enough, because the GPIO input is needed too (@angrykirc). Nobody had reported an ESP32 run by 2026-11-18. @_arcanum_ is collecting which RP2350 silicon revision people use.

Reported 8 GB
Source: @dalzyyy, 2026-11-11 (CMP 170hx Clean Room Discussion server)
Source: @scallion_9883, 2026-11-11 (CMP 170hx Clean Room Discussion server)
Source: @scallion_9883, 2026-11-11 (CMP 170hx Clean Room Discussion server)
Source: @dalzyyy, 2026-11-13 (CMP 170hx Clean Room Discussion server)
Source: @dalzyyy, 2026-11-13 (CMP 170hx Clean Room Discussion server)
Source: @_arcanum_, 2026-11-16
Source: @c133, 2026-11-17 (CMP 170hx Clean Room Discussion server)
Source: @samuel_hayden., 2026-11-17
Source: @samuel_hayden., 2026-11-17
Source: @angrykirc, 2026-11-17
Source: @angrykirc, 2026-11-17
Source: @samuel_hayden., 2026-11-17
Source: @con3po7016, 2026-11-18 (CMP 170hx Clean Room Discussion server)

Temperature, settings and attempts

These are per-card numbers, not a recipe. More: Voltage-glitch unlock (GA100VFI)#Attempts, reboots and time per card.

  • VRM temperature matters. @samuel_hayden.'s first card was only sensitive in a range around 40.5-42.5 C, read over I2C from the core voltage controller (MP2888B), not the GPU core, and the window is "going to be unique for your card and even amongst your own cards". One of @ichundes's two cards only glitched once it was up to temperature; @mbelli moved his VRM from 36 to 41 C between campaigns (the card cools while it is off the bus); @jonathanyjoerrington runs it cool with the fan at full, only right after boot.
  • Attempts run from tens to thousands: under 100 tries and usually under 2 minutes for @dalzyyy with the improved method; about 100 for @.shakdiesel; about 600 for @zanooda's first card while his second had not yet glitched; about 3000-4000 for each of @killred's cards; 652 without success for @mbelli and about 400 for @dr.bubotnic at the time. @killred reboots the host after each batch of 500 attempts, and when a run kept the same v_dip he told @zanooda "You need a reboot".
  • Lost reads and a dropped link do not always mean the dip was too low (@mbelli). @dalzyyy: try a single pulse instead of 3; the adaptive mode varies only v_dip; for a full sweep do 25 shots at each setting. Asked what recovery voltage range is safe to try, he answered that his notes put the best for his card at 625-650 mV, "so between 600-675mV maybe?".

Reported 8 GB
Source: @dalzyyy, 2026-11-13 (CMP 170hx Clean Room Discussion server)
Source: @dalzyyy, 2026-11-13 (CMP 170hx Clean Room Discussion server)
Source: @ichundes, 2026-11-17
Source: @jonathanyjoerrington, 2026-11-17
Source: @mbelli, 2026-11-17
Source: @.shakdiesel, 2026-11-17
Source: @mbelli, 2026-11-17
Source: @jonathanyjoerrington, 2026-11-17
Source: @mbelli, 2026-11-17
Source: @samuel_hayden., 2026-11-17
Source: @dr.bubotnic, 2026-11-17
Source: @samuel_hayden., 2026-11-17
Source: @samuel_hayden., 2026-11-17
Source: @samuel_hayden., 2026-11-17
Source: @dalzyyy, 2026-11-17
Source: @killred, 2026-11-18
Source: @zanooda, 2026-11-18
Source: @killred, 2026-11-18
Source: @killred, 2026-11-18
Source: @dalzyyy, 2026-11-18
Source: @killred, 2026-11-18
Source: @zanooda, 2026-11-18
Source: @dalzyyy, 2026-11-18
Source: @dalzyyy, 2026-11-18
Source: @dalzyyy, 2026-11-18

Before the first burn on a glitched card

Test in software first. Once DSOV is off, SW-IFF can test a change before you burn FF fields, and it is lost at a power cycle: "frankly it'd be better to use SW-IFF to test these things before burning FF fields" (@scallion_9883, 2026-11-13). @dalzyyy had SW-IFF working on a glitched 8 GB card on 2026-11-12, but the driver would not load yet. See also "Burn only what others have tested" under #Do not do these.

Reported 8 GB
Source: @dalzyyy, 2026-11-12 (CMP 170hx Clean Room Discussion server)
Source: @akuchen_66498, 2026-11-12 (CMP 170hx Clean Room Discussion server)
Source: @scallion_9883, 2026-11-13 (CMP 170hx Clean Room Discussion server)
Source: @scallion_9883, 2026-11-13 (CMP 170hx Clean Room Discussion server)
Source: @dalzyyy, 2026-11-13 (CMP 170hx Clean Room Discussion server)
Source: @nemz38_is_my_waifu_91412, 2026-11-16

Burn an inert canary first. @akuchen_66498 (2026-11-12): "Do not start with arbitrary unlock records. First use a semantically inert canary and verify its exact readback." Then burn one targeted record; afterwards re-sense or reset, load the stock driver, and check the effective status values, the CUDA SM count, a functional kernel and a GEMM. If records are already burned, debug (readback, re-sense or reset, the driver) before you burn more. In October he suggested a fuse dump before and after every write. A canary with bit 5 set ends the FF region (@stsoftware_88487 to the owner of the stuck card: "I think the problem is in bit5, in your 'canary'"; see #Do not do these).

Reported 8 GB
Source: @akuchen_66498, 2026-10-11 (CMP 170hx Clean Room Discussion server)
Source: @akuchen_66498, 2026-11-12 (CMP 170hx Clean Room Discussion server)
Source: @stsoftware_88487, 2026-11-16 (CMP 170hx Clean Room Discussion server)

Warning 8 GB Defective FBPs are the big brick risk. On the 8 GB card all re-fusing happens in the one glitch window, with no software test of a defective FBP first, and "If the fusing results in a unbootable card, there is no way to recover" (@petrikrohn_26497, 2026-11-11); that is why Kate unlocked only 80 GB. @mellojello_ is fairly sure a function-level reset can test records in the glitched state without touching the fuse macro (untested), and @replacedriver47 asked whether a dead FBPA could be tried for one boot through the software override instead (unanswered). Kate's view that FBPs can always be disabled again is on Fuses and floorsweep maps#Can FBPs marked defective be enabled and tested? (contested).

Source: @petrikrohn_26497, 2026-11-11 (CMP 170hx Clean Room Discussion server)
Source: @mellojello_, 2026-11-11 (CMP 170hx Clean Room Discussion server)
Source: @replacedriver47, 2026-11-18 (CMP 170hx Clean Room Discussion server)

Reported outcomes (short form)

Outcome reports are collected on Voltage-glitch unlock (GA100VFI)#Status and reported outcomes. In short: @dalzyyy had a glitched 8 GB card on 2026-11-12; by 2026-11-16 @mellojello_ counted about 15 replications, and @lia_oc said multiple people had replicated Kate's result. Finished cards include @zanooda's at 112 SMs with memtest passing and PCIe Gen4, @jonathanyjoerrington's covering 126 of 126 target SMs in a 60-second stress run, and two cards of @killred's, who now runs the step after the glitch automatically. Two results show what a burn did not fix: @.shakdiesel got 124 SMs but still 64 GB, and on one card cmpunlocker still added 4 SMs after the FF burn, because the GPC RECONFIG state is not in the records ("need clear RECONFIG for your GPCs", @dalzyyy); nobody has mapped that state to an FF record yet. On 2026-11-15 there was "no public recipe for the FF programming for legal reasons" (@zagitta); AI-written capture and planning tools followed the next day (see #Files and where they stop). @anon33anon asked owners to post full 512-row reads from before and after a burn to build a better calculator; the warning below against posting a full dump applies.

Proven 8 GB
Source: @dalzyyy, 2026-11-12 (CMP 170hx Clean Room Discussion server)
Source: @zagitta, 2026-11-15
Source: @mellojello_, 2026-11-16
Source: @lia_oc, 2026-11-16
Source: @.shakdiesel, 2026-11-17
Source: @.shakdiesel, 2026-11-17
Source: @.shakdiesel, 2026-11-17
Source: @dalzyyy, 2026-11-17
Source: @reala7096, 2026-11-17
Source: @dalzyyy, 2026-11-17
Source: @reala7096, 2026-11-17
Source: @jonathanyjoerrington, 2026-11-17
Source: @zanooda, 2026-11-17
Source: @zanooda, 2026-11-17
Source: @anon33anon, 2026-11-17
Source: @killred, 2026-11-18
Source: @killred, 2026-11-18

Do not do these

Warning Do not paste another card's FF words. ffcalc has also over-counted SM potential by looking at GPC defective state and missing TPC defects. A claim that several cards all hit 128 was called suspicious for that reason. Decode your own dump, then check the decoder against a register read.

Source: @killred, 2026-11-16
Source: @wakeupmbrc1, 2026-11-16

The claim behind that, in the order posted (2026-11-16). These are predictions before any unlock, not results.

  • @deejay9983: "4 of my 6 can hit 128 sm"; one card only reaches 64 GB, the other five clear 80 GB, none 96. Source: @deejay9983, 2026-11-16
    Source: @deejay9983, 2026-11-16
  • @wakeupmbrc1: "that looks suspicious, 128 are very rare"; the calculator may look only at the GPC defective state, since two of the other cards are exactly 16 SMs (one GPC) short. Earlier that day he had bought four cards in person after the seller let him bin them: 122, 122, 124 and 126 SMs, and the seller's four 64 GB cards included three at 128 SMs (his reading; the tool is not named). Source: @wakeupmbrc1, 2026-11-16
    Source: @wakeupmbrc1, 2026-11-16
    Source: @wakeupmbrc1, 2026-11-16
    Source: @wakeupmbrc1, 2026-11-16

Contested 8 GB

Warning Every card needs its own plan, and you need it before you glitch. "Daily reminder fuses are INDIVIDUAL per card, do not blindly burn from another card's plan" (@zagitta, 2026-11-18). When a member was about to glitch his first card, @zagitta asked whether he had a full plan for the fuse writes ("Otherwise glitching does you no good"), warned that there are many gotchas, and suggested posting the fuse dump and the writes you intend to do for review; "If you don't have a fuse burn plan glitching is of no use". On the clean room server @truvv0r asked for a posted list of writes to be taken down because "it sounds like recipe while its not", @shirogeek: "None of what's on discord should be taken as absolute truth and guarantee", and @kingsizexxxl agreed it is "definitely not the same on every card"; @alanoo concluded that a universal solution is far off. @lia_oc: burning fuses is at your own risk, and there is no universal tool yet. Build your plan from your own card's read with the tools under #Files and where they stop.

Source: @lia_oc, 2026-11-17
Source: @truvv0r, 2026-11-17 (CMP 170hx Clean Room Discussion server)
Source: @shirogeek, 2026-11-17 (CMP 170hx Clean Room Discussion server)
Source: @truvv0r, 2026-11-17 (CMP 170hx Clean Room Discussion server)
Source: @kingsizexxxl, 2026-11-17 (CMP 170hx Clean Room Discussion server)
Source: @alanoo, 2026-11-17 (CMP 170hx Clean Room Discussion server)
Source: @zagitta, 2026-11-18
Source: @zagitta, 2026-11-18
Source: @zagitta, 2026-11-18
Source: @zagitta, 2026-11-18
Source: @zagitta, 2026-11-18

Warning Do not invalidate factory records. One 8 GB card (the parameter set above) had its two factory chain-1 records invalidated by burning the chain field from 1 to 31, as a probe of how the window is parsed. The disable masks then latched to 0, which enables the physically dead FBPs and GPC. OTP cannot clear that bit. The parsed window on that card is rows 222 to 302 and it had no blank row left, so no replacement record can be appended. CRC was still 0x55. The array is not corrupt. The logic state is stuck. Appends outside 222 to 302 were ignored at boot. Row 303 bit 5, once set, blocked every later row.

Source: @port00, 2026-11-16
Source: @wgxli, 2026-11-16

Members who studied the parser give the same advice. @zarquon3157 (2026-11-17), after asking whether invalidated records make the border between FF and IFF movable: his point came from static analysis of software state machines, he has not seen hardware invalidate an FF record on his samples, and "I wouldn't go invalidating mid chain FF records until it is has been HW verified." @cute_sc2: writing an old record to all ones is only usable for records that are all 0 on a regular A100, not on a CMP.

Source: @zarquon3157, 2026-11-17 (CMP 170hx Clean Room Discussion server)
Source: @jacksmith7981, 2026-11-17 (CMP 170hx Clean Room Discussion server)
Source: @cute_sc2, 2026-11-17 (CMP 170hx Clean Room Discussion server)
Source: @zarquon3157, 2026-11-17 (CMP 170hx Clean Room Discussion server)

Update, 2026-11-18: the owner reports the card restored. The entry above (2026-11-16) says the stuck state cannot be cleared and that no replacement record can be appended. On 2026-11-17 @port00 gave the cause in his own words: the card "worked fine until invalidation" of the two factory records, which zeroed GPC_DISABLE and FBP_DISABLE, and the AI tool he used (mostly Qwen8Max and GLM5.3) had burned the bit-5 word into row 303 because "it didn't know about bit5", so the new records after it were never read. The card still showed in lspci but would not load the driver. @killred suggested making row 303 get discarded by burning its remaining bits to 1. On 2026-11-18 @port00 posted before-and-after screenshots, "the damage was revered", after burning row 303 to all ones ("convert binary to hex and you get 0xFFFF...."). @wakeupmbrc1 said he did not understand how the bit-5 row was invalidated, since rows 304 to 307 now read as FF writes. This is one owner's report with screenshots, not reproduced, and it stands beside the 2026-11-16 description above. Lessons posted in the thread: add row-invalidation checks to a plan even when the burn is fully planned (@wakeupmbrc1), and read Kate's blog before such a burn (@killred).

Reported 8 GB
Source: @killred, 2026-11-16
Source: @killred, 2026-11-16
Source: @port00, 2026-11-17
Source: @port00, 2026-11-17
Source: @wakeupmbrc1, 2026-11-17
Source: @wakeupmbrc1, 2026-11-17
Source: @port00, 2026-11-17
Source: @killred, 2026-11-17
Source: @killred, 2026-11-17
Source: @port00, 2026-11-17
Source: @port00, 2026-11-18
Source: @port00, 2026-11-18
Source: @wakeupmbrc1, 2026-11-18
Source: @wakeupmbrc1, 2026-11-18
Source: @port00, 2026-11-18
Source: @port00, 2026-11-18

Can a wrong burn be undone? Members disagree. In the order posted:

  • @dalzyyy (2026-11-16): "in terms of FF burning so long as you have rows left you can undo everything(dont take this as advice)". Source: @dalzyyy, 2026-11-16
  • @wakeupmbrc1: a fix still needs the Pico soldered on; @lia_oc: on the 8 GB card you need the Pico to burn at all. Source: @wakeupmbrc1, 2026-11-16
    Source: @lia_oc, 2026-11-16
  • @kingsizexxxl: there may be enough written by the FF blocks to keep the card from coming up. Source: @kingsizexxxl, 2026-11-16
  • @gnasherrr, on the stuck card: a wrong bit flashed into one row made the boot loader ignore all the following ones. Source: @gnasherrr, 2026-11-16
  • @samuel_hayden. (2026-11-17): a wrong value has to be corrected with more burns, "so you only have a limited number of attempts even though there are quite a few open". Source: @samuel_hayden., 2026-11-17
  • @wakeupmbrc1 (2026-11-17): "certains instructions in FF semantics can completely brick the card when written in IFF". Source: @wakeupmbrc1, 2026-11-17

Contested 8 GB

Warning Never burn a word with bit 5 set inside the FF region. A word with bit 5 set is the termination instruction that closes the FF pool: the IFF region starts after it, every FF record burned after it is ignored, and an OTP bit cannot be cleared. "i think bit5=1 writing should be strictly prohibited when doing any writing in FF" (@wakeupmbrc1, 2026-11-17; @port00 agreed), and it is his first example of things "you must NOT burn". That is what stuck the card above: the tool burned a bit-5 word into the row just after the card's last FF record, so the replacement records after it were never read. A bit-5 word further down would simply move the IFF start to that row. Check every word of a plan for bit 5 before you burn anything into FF rows, canaries included (@stsoftware_88487: "Bit5 - The Killer Canary").

Source: @stsoftware_88487, 2026-11-10 (CMP 170hx Clean Room Discussion server)
Source: @stsoftware_88487, 2026-11-16 (CMP 170hx Clean Room Discussion server)
Source: @stsoftware_88487, 2026-11-16 (CMP 170hx Clean Room Discussion server)
Source: @killred, 2026-11-17
Source: @port00, 2026-11-17
Source: @port00, 2026-11-17
Source: @wakeupmbrc1, 2026-11-17
Source: @wakeupmbrc1, 2026-11-17
Source: @wakeupmbrc1, 2026-11-17
Source: @wakeupmbrc1, 2026-11-17
Source: @port00, 2026-11-17
Source: @wakeupmbrc1, 2026-11-17
Source: @wakeupmbrc1, 2026-11-17
Source: @wakeupmbrc1, 2026-11-17

Warning Count your spare FF rows before you burn. Every correction, revert or re-disable costs a blank row, and a card with none left cannot take a fix. Counts differ per die: "off the top of my head 196" (@dalzyyy), "like 100-200 are empty in the valid range" (@killred, whose first free row for new records was 308), and about 38 zero rows after the last decoded record within rows 222-331 of one 8 GB dump (@akuchen_66498, 2026-10-06). An AI-generated answer that @petrikrohn_26497 posted, relayed by @68x0: "Some cards have a limited FF region that ends around fuse row 310. At least one card has already been bricked because it ran out of free FF rows. Before you burn anything, make sure you have plenty of spare FF rows." (AI-generated analysis; the stuck card above had no blank row left in its parsed window.) @simple98. found one of five newly bought cards with "almost fully written fuse data". @wakeupmbrc1 named too few free rows as the case where a wrong burn cannot be recovered. Another card's count is not yours: read your own.

Source: @akuchen_66498, 2026-10-06 (CMP 170hx Clean Room Discussion server)
Source: @dalzyyy, 2026-11-16
Source: @killred, 2026-11-16
Source: @wakeupmbrc1, 2026-11-16
Source: @killred, 2026-11-16
Source: @petrikrohn_26497, 2026-11-17 (CMP 170hx Clean Room Discussion server)
Source: @68x0, 2026-11-17
Source: @simple98., 2026-11-17

Warning Do not write the main array through the RIR tool. RWL=1 is the RIR bank. Main-array FF writes use RWL=0 and only blank rows below the IFF boundary. A writer reviewed from the 2026-11-16 posts never touches DEBUGCTRL or RWL, refuses a row that is not blank, and read-back verifies. The planner beside it is hardcoded for the 8 GB id 0x20c2 and will not plan a 0x2082 card. Add nothing to a burn until the reversible test has already survived a driver bind.

Source: @darkmaniac7, 2026-11-17 (lab notes)

Warning Do not trust a generated burn plan that you have not checked against the card. Two members were talked down from running plans where models contradicted each other. One was told a model had already fried a card once. Build the check offline, compare it to your dump, and stop if it wants to enable a defective unit or to rewrite a factory record.

Source: @jonathanyjoerrington, 2026-11-16
Source: @zagitta, 2026-11-16
Source: @kingsizexxxl, 2026-11-16

@galygious posted the checking steps on 2026-11-15, in English and Russian: give the model Kate's blog, the repository and your own card's dump; have it build an offline decoder and checker first, so you can inspect the proposed records, confirm they keep unrelated settings and the defect masks, and verify that the destination rows are unused; keep anything undocumented explicitly unknown. "Two models agreeing isn’t verification, the values need to check out against the sources and your dump."

Source: @galygious, 2026-11-15

Warning Check the exact 32-bit word your tool will write. On two GA102 CMP 90HX cards (2026-11-03) @truvv0r tried to plant a two-word IFF instruction in unused OTP rows. On both cards the header word landed as intended, but the data word meant to be 1 landed with bit 27 set instead of bit 0 (0x08000000). A read-only check afterwards showed the override still disabled: the rows never executed, and they are spent. @n2dcrypt relayed an AI reading that the tool wrote the field mask instead of the value; @truvv0r later said the only reason the attempt failed was a bug in his Python script that set an incorrect value. Before any program command, dump the exact words the tool will write and check them against an independent decode.

Source: @truvv0r, 2026-11-03 (CMP 170hx Clean Room Discussion server)
Source: @truvv0r, 2026-11-03 (CMP 170hx Clean Room Discussion server)
Source: @n2dcrypt, 2026-11-03 (CMP 170hx Clean Room Discussion server)
Source: @truvv0r, 2026-11-16 (CMP 170hx Clean Room Discussion server)

Warning Burn only what others have tested, and test it in software first. @petrikrohn_26497 (2026-11-17): wait a few days before burning and bricking a card; "Only do burns that have been tested and confirmed by others." If you try something new, read all the public material, understand it, and ask before making mistakes. "Before testing any new live burns on 8GB card, make sure they have first been tested with software overrides (SW-IFF) on 10GB cards." He wrote it after @kingsizexxxl said the community was being "a bit overcautious" and offered a list of the burns he knows. @cmdrwaka324: a script that calculates a plan can be generic, a fixed list of writes cannot. On 2026-11-09 @n2dcrypt held back from more burns because they cannot be undone and cards had already been reported bricked.

Source: @n2dcrypt, 2026-11-09 (CMP 170hx Clean Room Discussion server)
Source: @kingsizexxxl, 2026-11-17 (CMP 170hx Clean Room Discussion server)
Source: @petrikrohn_26497, 2026-11-17 (CMP 170hx Clean Room Discussion server)
Source: @cmdrwaka324, 2026-11-17 (CMP 170hx Clean Room Discussion server)

Warning Do not post or reuse a full FF dump. A full dump can carry per-card key material. Publish the disable and defect masks you used, not the raw array.

Source: @galygious, 2026-11-16

Warning Do not flash an A100 VBIOS onto a board with a small MMIO window. Above 4G has to be on, CSM off, Resizable BAR on, on the order of 128 GB of BAR per card. A Z790 hung at POST D4 or D5. An MSI Z790 had no option that booted it. x870e and one X12 Xeon board were reported to work. If the card still shows up in lspci you can usually recover. You cannot write a card that does not boot. "PCIe magic bits" were called out as a way to brick one for good.

Source: @kingsizexxxl, 2026-11-16
Source: @._.patrik, 2026-11-16
Source: @dalzyyy, 2026-11-16
Source: @matstatman, 2026-11-16

More from 2026-11-16 and 11-17: the A100 80 GB PCIe VBIOS "needs big bar size apparently" (@lia_oc), "like a100 needs like 128gb bar" (@kingsizexxxl). System RAM does not limit the BAR size: @cute_sc2 runs a 64 GB card with the matching BAR on an X99 host with 16 GB of RAM, as long as the board supports UEFI and large BARs. @mellojello_ suggested disabling Resizable BAR and letting Linux reassign the resources (untested). The VBIOS does not care about the operating system, "it only cares about your BAR size and your motherboard support" (@samuel_hayden.). Where to get the 80 GB PCIe image is open: per @beshox., TechPowerUp's collection has only the SXM image and the 40 GB PCIe one. Whether a changed device ID can be reversed was asked and not answered (@dr.bubotnic).

Source: @lia_oc, 2026-11-16
Source: @dr.bubotnic, 2026-11-16
Source: @cute_sc2, 2026-11-16
Source: @kingsizexxxl, 2026-11-16
Source: @mellojello_, 2026-11-16
Source: @beshox., 2026-11-17
Source: @beshox., 2026-11-17
Source: @samuel_hayden., 2026-11-17
Source: @samuel_hayden., 2026-11-17

Warning Pads come off the board. One member tore four pads. Another lost a 2.21k resistor (R1007). The heatsink is a lot of thermal mass. Microscope, hot air, and a donor card to practise on, before the card you care about. A short or loose PCIe cable also imitates a dead card. Swap the cable before you decide the silicon is gone.

Source: @deejay9983, 2026-11-16
Source: @tekkenz, 2026-11-16
Source: @voidgiraffe, 2026-11-16

Warning A fuse flag can lock the RIR records for good. @bigptoughneigh (2026-11-06): "there is a flag in the fuse array that disables RIR modification, even on unused rir slots. Ask me how I know". He did not name the flag, the card variant or what set it. A free RIR slot is not a guarantee that a gate burn will land: read the RIR state and the whole fuse array first, and do not burn anything whose effect on RIR programming you do not know.

Source: @bigptoughneigh, 2026-11-06 (CMP 170hx Clean Room Discussion server)

Warning Program as few single rows as you can, read each back, and stop after a failure. @truvv0r (2026-10-29): "a session of ~40 program cycles in 2 hours on two cards permanently killed the fuse WRITE path on both (reads unaffected, survives power cycles)"; his rule then was never to batch several write commands under one programming-voltage window and never to repeat program cycles back to back. On 2026-11-02, challenged by @stsoftware_88487, he corrected it: the cards were GA102 CMP 90HX, not GA100, and he had not measured the programming rail or proven permanent damage, only that writes stopped succeeding after many attempts; one successful single-row probe is no permission to run a ten-row set. On 2026-11-16 he said the only reason his attempt failed was a bug in his script. The cause is unproven; the safe reading stands.

Source: @truvv0r, 2026-10-29 (CMP 170hx Clean Room Discussion server)
Source: @truvv0r, 2026-10-29 (CMP 170hx Clean Room Discussion server)
Source: @stsoftware_88487, 2026-11-02 (CMP 170hx Clean Room Discussion server)
Source: @truvv0r, 2026-11-02 (CMP 170hx Clean Room Discussion server)
Source: @truvv0r, 2026-11-16 (CMP 170hx Clean Room Discussion server)

Warning Do not keep firing "revival" burns at a card that will not boot. @stsoftware_88487 (2026-11-17): his card with a bit-5 error did not boot ("GFW_BOOT progress 0x1"), and after several tries of "revival" fuses, written as correct IFF opcodes, it "went into the sunset": it no longer shows on the PCIe bus. His own investigation points to an overflow of the VBIOS black-box journal (BBX) region, and he thinks the dead VBIOS state happened while old journal content was being erased at the moment a power glitch was applied. Nobody else confirmed the cause. @port00 asked others to review his repair plan before burning it.

Source: @port00, 2026-11-17 (CMP 170hx Clean Room Discussion server)
Source: @stsoftware_88487, 2026-11-17 (CMP 170hx Clean Room Discussion server)
Source: @stsoftware_88487, 2026-11-17 (CMP 170hx Clean Room Discussion server)

Warning Research has already cost cards: test on cheaper hardware and reproduce first. Asked in October how many cards were lost along the way, @bigptoughneigh counted fewer than 10 for his group (one of them his own), while another member's crew had lost dozens on their own. @cute_sc2 had killed one the day before: "It was user error." His advice to a newcomer: test on less expensive hardware first, try to reproduce other people's findings, and "Also expect that you might destroy cards due to mistakes."

Source: @akuchen_66498, 2026-10-03 (CMP 170hx Clean Room Discussion server)
Source: @bigptoughneigh, 2026-10-03 (CMP 170hx Clean Room Discussion server)
Source: @cute_sc2, 2026-10-03 (CMP 170hx Clean Room Discussion server)
Source: @cute_sc2, 2026-10-03 (CMP 170hx Clean Room Discussion server)
Source: @cute_sc2, 2026-10-03 (CMP 170hx Clean Room Discussion server)
Source: @cute_sc2, 2026-10-03 (CMP 170hx Clean Room Discussion server)

Reported Do not burn ECC_EN_CONTROL. An AI-found table that @68x0 posted on 2026-11-17 puts ECC_EN at row 6 bit 28 (copy at row 7 bit 28) and ECC_EN_CONTROL at row 6 bit 29 (copy at row 7 bit 29), FF chain not established. @truvv0r answered the ECC_EN_CONTROL line with "Never do this" and gave no reason. In October @mountains1605 had found one 8 GB die among nine shared dumps with those two bits burned to 1 while the live ECC option read 0; his AI-assisted explanation, a factory correction record, was questioned by @mbelli's AI review. A burned ECC fuse alone does not give working ECC (see #What is not a fuse flip).

Source: @mountains1605, 2026-10-03 (CMP 170hx Clean Room Discussion server)
Source: @mbelli, 2026-10-03 (CMP 170hx Clean Room Discussion server)
Source: @68x0, 2026-11-17 (CMP 170hx Clean Room Discussion server)
Source: @truvv0r, 2026-11-17 (CMP 170hx Clean Room Discussion server)

Theory Driver updates and the firmware revocation fuses. @zarquon3157 (2026-10-26) wrote that the fuse array carries version-revocation fields for the SEC2, PMU and GSP firmware; he marked part of the analysis as AI commentary he was less sure of. @truvv0r's conclusion: "So drivers updates should be done with extra caution". No member has reported a driver update that burned such a fuse on a CMP card (AI-assisted analysis, not measured). Before moving an unlocked card to a new driver branch, read the reports on Driver unlock first.

Source: @zarquon3157, 2026-10-26 (CMP 170hx Clean Room Discussion server)
Source: @truvv0r, 2026-10-26 (CMP 170hx Clean Room Discussion server)

Theory 10 GB Re-running devinit at runtime can over-volt the card. @bigptoughneigh (2026-07-27), working on the 10 GB card's 80 GB geometry: "Re-executing devinit via PMU at runtime is quite risky"; the region that holds memory timing and MRS programming is part of the training section that also sets clocks, PLLs and VID-PWM, and he thinks a wrong value could push the VRM past 1.3 V. An AI analysis he pasted (AI-generated analysis, not measured; he was not sure it was right) called runtime timing changes a dead end: derived strap timings at the 80 GB geometry failed driver init, and per-strap timings only work if set at boot. No damage was reported.

Source: @bigptoughneigh, 2026-07-27 (CMP 170hx Clean Room Discussion server)
Source: @bigptoughneigh, 2026-07-27 (CMP 170hx Clean Room Discussion server)
Source: @bigptoughneigh, 2026-07-27 (CMP 170hx Clean Room Discussion server)

What is not a fuse flip

Tested with SW-IFF on a gate-burned 10 GB card, then the driver bound again. All of it reverted.

  • Gen3. Clearing OPT_DISABLE_GEN3_SPEED (0x820250) and the magic-bit register made Link Cap2 advertise Gen3 and Gen4. The primary link cap stayed at Gen2, and a forced retrain dropped the link. Recovery was a power cycle. The build that worked kept cmpunlocker's Gen2 booter scaffolding, swapped in the Gen3 GSP block, and overrode OPT_PCIE_BOOT_GEN3_DIS (0x820580), first for one card, then, with cmp_fuse_bdf left out, for both. The 108-SM card then trained to Gen3 x4 on the first attempt, 0 Xid, 0 AER. After the retrain published in lixiangwuxian/170hx_tips, both ends read 8 GT/s x4, no new AER, no new Xid. Pinned copies went from 1.67 / 1.71 GB/s (Gen2 x4) to 3.25 / 3.38 GB/s (Gen3 x4). A 30 minute soak stayed at Gen3 x4, peak 65 C. A second 30 minute soak with both cards loaded at once: 120 of 120 samples at Gen3 x4 on each, 0 AER, 0 Xid, 3.24 / 3.38 GB/s each. Width stayed x4. x16 is still the PCIe x16 capacitor mod. The link is volatile and has to be retrained every boot. The guest cannot setpci the host root port, so this retrain does not run inside the VM. "Booter Load failed" lines also appear on the working Gen2 module. They are not the failure signal. Link Cap2 staying at 2.5 GT/s is.
  • Gen4. Clearing bit 25 of PCIE_MAGIC_BITS_D (0x820520) only adds 16 GT/s to Link Cap2. Training it also takes the Gen4 PHY init and one config-space retrain from a link already at Gen3; done that way, both cards ran 16 GT/s x4 at both ends with 0 AER and 0 Xid, and 6.37 / 6.76 GB/s pinned copies. The other Gen4 bits of that register stayed set. All of it is volatile and redone at every boot: burn nothing for it, and make one retrain attempt per boot, not a loop. See PCIe Gen4 on the 10 GB card. Proven 10 GB Source: @darkmaniac7, 2026-11-17 (lab notes)
  • ECC. The ECC fuse bit applied in hardware and nvidia-smi still showed ECC mode N/A. It needs the driver patch. Do not burn it first.
  • NVLink. The disable fuse cleared and the defect fuse was already 0. The driver still reported no NVLink. You still need the driver, the PHY, and the missing board parts. A single 3.3 V jumper to the middle bridge connector is the minimal hardware path members report as having trained. See NVLink. An SM-only burn does not turn it on.
  • V/F offset on the 10 GB VBIOS 92.00.66.00.02. NVML reports the GPC range as [0..0]. An in-memory override of the card's own image, not a flash, opened [-1000..+1000]. At a 250 W cap, with the case fans actually spinning, a +190 offset on the 122 SM card held about 1172 MHz and about 8 to 9 percent more throughput than offset 0, 0 Xid across the ladder up to +220. The clock flattened after +190. Fix airflow before you judge a number: a passive card in a case whose fan curve follows CPU temperature sat at 85 C and thermal-throttled until the fans were run at 12 V. Do not flash this patch. It is one VBIOS revision, applied through the driver's in-memory override.

Reported 10 GB
Source: @darkmaniac7, 2026-11-17 (lab notes)
Source: repo github.com/lixiangwuxian/170hx_tips by lixiangwuxian on GitHub

Members report one identity change that is a fuse burn. @petrikrohn_26497 (2026-11-11, passing on news): "It is possible to re-fuse the 10GB card with the device ID of the A100 PCIe 40GB and flash it with a A100 VBIOS." On 2026-11-14 @zagitta reported that he flashed an A100 VBIOS after changing the device ID and the card was "fully unlocked without any software" (screenshot), with about 5-10 W more at idle with all 128 SMs. He then qualified it: the SS0 and SS1 issue-rate limits still had to be written through IFF, and the FP64 speed bit sits in the regular fuses; @dalzyyy: FP64 and ECC poison containment are column-backed and cannot be re-enabled by FF records. @petrikrohn_26497 had understood SS0 and SS1 to come from the VBIOS; that is open. The device-ID change is a permanent burn, and without it an A100 VBIOS leaves the card unbootable; mind the BAR warning above. More: Voltage-glitch unlock (GA100VFI)#Re-fusing the device ID, then an A100 VBIOS.

Reported 8 GB and 10 GB
Source: @petrikrohn_26497, 2026-11-11 (CMP 170hx Clean Room Discussion server)
Source: @zagitta, 2026-11-14 (CMP 170hx Clean Room Discussion server)
Source: @zagitta, 2026-11-14 (CMP 170hx Clean Room Discussion server)
Source: @zagitta, 2026-11-14 (CMP 170hx Clean Room Discussion server)
Source: @dalzyyy, 2026-11-15 (CMP 170hx Clean Room Discussion server)
Source: @zagitta, 2026-11-15 (CMP 170hx Clean Room Discussion server)
Source: @zagitta, 2026-11-16 (CMP 170hx Clean Room Discussion server)
Source: @petrikrohn_26497, 2026-11-16 (CMP 170hx Clean Room Discussion server)
Source: @zagitta, 2026-11-16 (CMP 170hx Clean Room Discussion server)

Files and where they stop

  • Gate burn: lixiangwuxian/170hx-rir-tool. RIR rows only.
  • Glitcher and option maps: kate-land/ga100vfi. Hardware, 8 GB path.
  • FF calculator: sayhellotojungle/cmp170hx-ffcalc. It can over-count SMs. Check it.
  • Driver: amoghmunikote/cmpunlocker, and the Gen3 notes in lixiangwuxian/170hx_tips.
  • One-card research archive, already redacted for this site: CMPGlitch fuse research archive.
  • A 16-source reading list posted with those tools: Kate's blog, the repos above, patent US9230678B2, NVFlash, the open GPU kernel modules, driver 610.43.02, the A100 80 GB brief, and the JRex286 GA100 register notes. Source: @kingsizexxxl, 2026-11-16
  • Reported Glitch controller for a Raspberry Pi Pico 2: dalzyu/pico-ga100vfi by @dalzyyy. It does the voltage-fault step only; it resets the GPU through PEX_RST so fuse sensing can be targeted again, while the I2C link sets the voltage dip (@mbelli). You still program your own FF records afterwards: "this repo get you to the unlock then you need to program the specific fuseless fuses" (@galygious). Pull the latest commit: its author fixed the adaptive search on 2026-11-17, and it still varies only v_dip. RP2040 alternatives are under "Controller boards" above. Source: @galygious, 2026-11-15
    Source: @mbelli, 2026-11-15
    Source: @galygious, 2026-11-15
    Source: @dalzyyy, 2026-11-17
    Source: @dalzyyy, 2026-11-17
    Source: @dalzyyy, 2026-11-18
  • Reported Fuse console: portz3r0/pico-ga100vfi-console by @port00, a live view of a fuse dump for the Pico glitcher ("still improving"). Its author's card is the stuck and restored card above: one person's tool, tested on his own card. Source: @port00, 2026-11-16
    Source: @port00, 2026-11-17
    Source: @port00, 2026-11-17 (CMP 170hx Clean Room Discussion server)
    Source: @port00, 2026-11-17 (CMP 170hx Clean Room Discussion server)
  • Reported FF capture and planner: kprinssu/ga100-ff-tools by @68x0, released 2026-11-17 and written by an AI model ("This is slop, but the dumper does seem to be reading my FF rows"), made more generic from the CMPGlitch work. Its author spot-checked the output, asked for help verifying the chains and rows, and said the model had trouble with the Gen4 bits. An earlier zip of the same scripts went round on 2026-11-16. No member had reported a burn made with it by 2026-11-18. Check any plan it writes against your own read. Source: @68x0, 2026-11-15
    Source: @68x0, 2026-11-15
    Source: @68x0, 2026-11-16
    Source: @68x0, 2026-11-16
    Source: @68x0, 2026-11-16
    Source: @68x0, 2026-11-16 (CMP 170hx Clean Room Discussion server)
    Source: @68x0, 2026-11-16
    Source: @68x0, 2026-11-17
    Source: @68x0, 2026-11-17
  • Reported A per-card planner by @killred, written with an AI model, worked on his cards ("It gonna be automated for all of the cards"); he will release it after testing it on all of them, and it was not public on 2026-11-18. Asked whether the mod could be done in software, @killred said no, and @wakeupmbrc1 said people have explored every single bit. Source: @killred, 2026-11-18
    Source: @killred, 2026-11-18
    Source: @killred, 2026-11-18
    Source: @killred, 2026-11-18
    Source: @wakeupmbrc1, 2026-11-18

Raw FF dumps and per-card burn scripts are not hosted here. A dump can contain key material, and a script that ran on one die will enable dead units on the next one. The public writers that do exist still have to be read against your card: the 8 GB planner refuses 0x2082, and a writer is only as safe as the plan you feed it.

Related