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.

Voltage-glitch unlock: lab results and fuse findings

From CMP Wiki

On 18 and 19 November 2026 the lab took four 8 GB cards (10de:20c2) from its eight-card server to a bench and ran the Pico glitch on them; the setup, the tool fixes and the burn flow are on Voltage-glitch unlock: lab method on a Pico 2 W. Two cards are finished: compute from 70 to 122 and 120 SMs (126 and 124 with the RECONFIG override), the 80 GB memory layout in the fuses, and PCIe fuses that advertise 2.5 to 16 GT/s, all read after a cold boot with the planner reporting nothing left to burn. A third card never reached the glitch window in about 4,000 shots. The fourth has no core rail and needs a board repair.

The FF records for the memory and PCIe sets were the same on both finished cards, so they are given below with the block values they replace. The compute records differ per card and are not given: build yours from your own read.

Proven 8 GB
Source: @darkmaniac7, 2026-11-19 (lab notes)

Warning A record replaces a whole 16-bit block. The records on this page are right only for a card whose blocks hold the same old values as the lab's cards. Read your own card first; if any old value differs, your record differs, and the planner has to build it. FF burns cannot be undone: Unlock without bricking.

Results

Card A Card B Card C Card D
SM ceiling from the fuse census (defects kept) 126 124 124 128
Before 70 SMs at the fuse floor, 74 with the override, 64 GB same same did not enumerate
Glitch goldilocks in 11 shots; a second session for memory and PCIe took about 1,200 shots goldilocks on the first cell at 356 mV; a short second session for one PCIe record no goldilocks in about 4,000 shots not glitched
FF records burned 14 (6 compute, 5 memory, 3 PCIe) 16 (7 compute, 5 memory, 3 PCIe, then the bit-9 record) none none
SMs after a cold boot 122, 126 with the RECONFIG override 120, 124 with the override stock
Memory fuses after a cold boot FBP disable 0x840, FBIO, FBPA and ROP_L2 disable 0xC03000: 10 of 12 FBPs, the 80 GB layout; HBM trained clean at boot same stock
PCIe fuses after a cold boot PCIE_MAGIC_BITS_D 0x14680000, OPT_PCIE_BOOT_GEN3_DIS 0, OPT_DISABLE_GEN3_SPEED 0; LnkCap2 lists 2.5-16 GT/s same stock
Planner on the new read NO-OP: 85 records in rows 222-306 all parsed, 181 blank FF rows left NO-OP: 88 records in rows 222-309 all parsed, before the bit-9 record

Checks after each cold boot: GFW 0xFF, CRC 0x55, DSOV back to 1 (normal for a cold card), two 512-row reads matching. Both finished cards already had the PCIe x16 capacitor mod.

Card B: the session that burned its compute, memory and PCIe records, and what the cold boot read back.

Proven 8 GB
Source: @darkmaniac7, 2026-11-19 (lab notes)

Still to do on the driver side

The fuses are done; using them is driver work, not yet run on these two cards on 2026-11-19:

  • 80 GB: the mainstream unlocker ignores the extra 16 GB; it needs a cmpunlocker memory-geometry patch for the 80 GB layout. Kate Land reports 80 GiB on nine of ten burned 8 GB cards (her results), and one member's burned card stayed at 64 GB (members' counts). Prove it with a workload that writes and reads back above the old limit (Verifying an unlock).
  • Gen3 and Gen4: LnkCap2 advertising 16 GT/s does not train the link. Expect the driver route proven on the 10 GB card: the Gen3 retrain, the Gen4 PHY init and a Gen4 retrain (PCIe Gen4 on the 10 GB card). On the bench riser the link stayed Gen1 x8.

Reported 8 GB
Source: @darkmaniac7, 2026-11-19 (lab notes)

Fuse findings

FF record format, as on Fuses and floorsweep maps: one 32-bit word, data[31:16] | block[15:6] | chain[4:0], bit 5 clear. The word carries the new value of the whole 16-bit block, and the last record for a block wins. For example 0x058209C2 is chain 2, block 39, new block value 0x0582.

OPT_DISABLE_GEN3_SPEED is chain 2 block 39 bit 9

Which bit of chain 2 block 39 holds the Gen3 speed disable was open: both bit 1 and bit 9 are set at stock (block value 0x0782), and cmp170hx-ffcalc left the two as unresolved candidates.

  • On card B the lab burned the bit-1 record first (0x078009C2, block 39 0x0782 to 0x0780). After a cold boot OPT_DISABLE_GEN3_SPEED (0x820250) still read 1.
  • A second record that clears bit 9 and puts bit 1 back (0x058209C2, 0x0782 to 0x0582) went in during a short second glitch session. After a cold boot 0x820250 read 0.
  • On card A the bit-9 record went in directly, and 0x820250 read 0 after the cold boot.

Bit 9 is the Gen3 speed disable; bit 1 does nothing for it. Card B's bit-1 record is a spent row but harmless, because the later record for the block wins. A planner should clear bit 9 alone and leave bit 1 set.

Proven 8 GB Measured on two cards, each after a cold boot.
Source: @darkmaniac7, 2026-11-19 (lab notes)

The PCIe set: LnkCap2 to 2.5-16 GT/s

Three chain-2 records, the same on both cards:

Register Bit in the FF chain Old block New block Record Register after a cold boot
PCIE_MAGIC_BITS_D (0x820520), bit 25 (the Gen4 disable) block 47 bit 3 0x0059 0x0051 0x00510BC2 0x16680000 to 0x14680000
OPT_PCIE_BOOT_GEN3_DIS (0x820580) block 49 bit 7 0x00E4 0x0064 0x00640C42 1 to 0
OPT_DISABLE_GEN3_SPEED (0x820250) block 39 bit 9 0x0782 0x0582 0x058209C2 1 to 0

With all three, LnkCap2 lists 2.5, 5, 8 and 16 GT/s after a cold boot; before, the same bench read 2.5 GT/s only. Not burned: OPT_PCIE_BOOT_GEN23_DIS (0x82057C) and the A100 values for the other PCIE_MAGIC_BITS_D bits. This is the same set of three fields that gave the lab's 10 GB cards Gen4 as a volatile driver override (PCIe Gen4 on the 10 GB card); on the 8 GB cards it is now permanent.

Proven 8 GB
Source: @darkmaniac7, 2026-11-19 (lab notes)

The memory set: the 80 GB layout

Five chain-1 records with the FBP record last. Both cards had the same old values in these blocks, so the records were the same:

Block Field Old block New block Record
5 FBIO disable 0x9867 0x8007 0x80070141
6 FBPA disable, first part 0x6601 0x0601 0x06010181
7 FBPA disable, second part 0x0198 0x0180 0x018001C1
20 ROP_L2 disable 0x1986 0x1800 0x18000501
19 (last) FBP disable 0x4290 0x4200 0x420004C1

After a cold boot: FBP disable 0x852 to 0x840, and the FBIO and ROP_L2 disable registers 0xC0330C to 0xC03000 (the FBPA mask decodes to 0xC03000 from the FF rows; the lab's snapshot tool has no FBPA register read). FBPs 1 and 4, disabled without a defect flag, are back on. FBPs 6 and 11, the two marked defective on every card in the lab's census, stay off: 10 of 12 FBPs, the 80 GB layout, and the HBM trained clean at boot. 0xC03000 is also the 10 GB card's stock FBPA mask. Most 8 GB cards read so far share this pattern, but at least one member's card has different FBPs marked defective (Floorsweep read of the 8 GB card: gated versus defective partitions): on such a card these records are wrong.

Proven 8 GB Proven for the fuse layout and boot training on two cards. Running 80 GB under a driver was not done yet (above).
Source: @darkmaniac7, 2026-11-19 (lab notes)

The compute set

Per card, built from the card's own read: the TPC disable masks in chain 1 blocks 11 to 15, PES disable in blocks 15 to 17, and the GPC disable (block 8) last. Card A took 6 records and card B 7. The GPC disable went to 0 on both, and the TPC defect marks stayed: one TPC in GPC 0 on card A (122 SMs), one each in GPC 5 and GPC 7 on card B (120 SMs). The last 4 SMs sit behind the RECONFIG register and come back with the driver's override (126 and 124), as on the 10 GB card (RECONFIG). After the compute burn on card A the boot parser read the FF rows at least to row 298 on that die.

Proven 8 GB
Source: @darkmaniac7, 2026-11-19 (lab notes)

Fuse rows in the goldilocks state

In the goldilocks state rows 0-487 read the same as in the cold read, but the IFF rows 488-507 read differently (a cold read shows the fixed stock pattern). Do not take that as corruption; compare rows 0-487 before you burn.

Reported 8 GB
Source: @darkmaniac7, 2026-11-19 (lab notes)

The card that would not glitch (card C): unsolved

Card C (census ceiling 124, the same voltage-controller profile as card A) never gave a real DSOV 0 in about 4,000 shots over four sessions. It went back to stock and healthy after every session. What was measured, for anyone with a card like it:

  • Never "over". Every valid shot read DSOV 1: the card goes from under straight to a crash. Two shots logged as DSOV 0 were poisoned shots with nothing read (why).
  • Depth. At 40 C, 375 mV was under and 368.75 mV lost, with no code in between. At 369 mV, pulses of 80-126 µs were about 60 percent under, 25 percent lost and 10 percent NACK; 128 µs and longer were all lost. Cool (33-36 C) moved the edge up one step, 43.7 C moved it down, and at 47-49 C every shot was lost.
  • A timing knife edge. Single 40 µs dips at 332-356 mV did no harm anywhere from t_dip -180 to +15 µs. At +17 µs 0 to 2 of 10 were lost, at +18 about two thirds, at +19 8 or 9 of 10: a 1 to 2 µs cliff, the same at every depth. Only at 338-344 mV and t_dip +20 to +60 µs did about 1 shot in 8 come back poisoned, the only sign that a dip reached the fuse read.
  • The standard train dies on its first pulse's tail. The standard 3 x 198 µs train survived at t_dip -131 µs and earlier (30 of 30 at 356 mV) and was lost at -130 µs and later (11 of 12 at -130 µs, every shot from there to -60 µs). A single 198 µs pulse showed the same edge. The edge did not drift over 360 shots and did not move between 32 and 39 C.
  • First-shot mode moves the edge, not the outcome. With a park before every shot the card mostly survived much deeper dips (rail troughs near 205 mV on the ADC) and the lethal edge moved about 45 µs later, to t_dip -90 to -80 µs. Still no flip and no poisoned shot in about 230 shots.
  • Strictly binary. In the edge zone, in both modes, every shot was either under or lost. The lab's reading: on this card the fuse read and the event that kills the boot fall within the same few microseconds, and a voltage command over PMBus cannot separate them.
NVVDD on card C at four settings, read by the Pico ADC (debug firmware). The dashed line is the commanded dip; under and lost shots look the same on the rail.
  • NVVDD traces (Pico ADC, debug firmware): the dips land 35-40 mV below the commanded value, each recovery edge overshoots by about 130 mV (850-870 mV on the ADC), and under and lost shots look the same on the rail. Card C dies at the rail depth, about 0.32 V on the ADC, at which the tool author's card reaches goldilocks. The lethal windows: the rail low in about the last 30 µs before the PEX_RST release, and again around +730 to +830 µs. Dips elsewhere did no harm, even at 306 mV.
  • Did not help: recovery voltages of 600-700 mV, 2-pulse trains, a 3 s wait before reading, cooling, heating, and narrow 40 µs pulses placed just before the edge.
  • Not tried yet: the original PWM-VID method from Kate Land's write-up (remove the 0402 0 ohm link between the GPU's PWM-VID output and the MP2888B, drive the PWM from your own board and time the glitch off the NVVDD power-up ramp; see Kate's hookup), a firmware change that holds SCL low after the dip so the GPU's devinit cannot talk to the controller, and a finer pre-release timing map with 36-50 µs pulses.

Reported 8 GB Unsolved. One card, one host.
Source: @darkmaniac7, 2026-11-19 (lab notes)
Source: repo github.com/dalzyu/pico-ga100vfi by dalzyu on GitHub
Source: repo github.com/kate-land/ga100vfi by kate-land on GitHub, 2026-11-10

A card with no core rail (card D)

Card D: the four soldered joints of its 12 V input pigtail. 12 V was present here, but NVVDD stayed at 0.016 V.
  • Symptom: never enumerated, in two different risers and directly in the slot, after several reseats; another card in the same slot came up. It had been running in the server the day before. It was never glitched.
  • I2C: the MP2888B did not answer at 0x25: NACK with the probes off, and I2C timeouts with them on (a line held low in the middle of a transfer). A working card answered on the same leads. The I2C pull-ups read 3.3 V and PEX_RST_BUF* 1.8 V, so the slot's logic supplies were up.
  • Power: 12 V present at the card's power input. Powered off, NVVDD to ground read 1.8 ohm (3.3 ohm on a healthy card; not a short) and 12 V to ground above 200 ohm and climbing, as on the healthy card. Powered on, NVVDD read 0.016 V: the core rail never starts.
  • Reading: the core-voltage controller is not running: no supply, no enable, or damage, most likely from handling when the card was pulled. A board repair, not a glitch problem. Next steps: compare photos of the back with a working card for a missing or cracked part, then inspect the power stages and the controller with the heatsink off. See Recovery#Dead or damaged cards: board-level repair.

Reported 8 GB
Source: @darkmaniac7, 2026-11-19 (lab notes)

Tools

The glitch-side tools behind these results are public at darkmaniac7/cmp170hx-8gb-vfi-tools: the patched firmware, park and rescue, the register snapshot, NVVDD trace capture, the stop-and-hold hunter and the edge tracker used on card C. The burn step is not part of it. What each tool does: method page.

Reported 8 GB
Source: repo github.com/darkmaniac7/cmp170hx-8gb-vfi-tools by darkmaniac7 on GitHub
Source: @darkmaniac7, 2026-11-19 (lab notes)

See also