Voltage Glitching Unlocks Full A100 Capabilities on NVIDIA CMP 170HX Mining Cards

TL;DR

A voltage‑fault‑injection (VFI) exploit on the GA100 fuse controller can blow arbitrary fuseless fuses, unlocking all SMs, full HBM capacity, PCIe Gen4, NVLink and other datacenter features on a CMP 170HX mining card – effectively turning it into an A100‑80G that runs ~1.2× faster than a stock A100.


What was discovered?

  • The GA100 die contains 16 384 eFuses organized into regular, fuseless (FF), IFF and RIR regions.
  • "Fuseless" fuses encode option‑words that enable or disable major blocks (SMs, FBPs, NVLink, PCIe speed, etc.).
  • Blowing an FF fuse does not physically erase a fuse; it writes a new 16‑bit block that overrides earlier values.
  • By using VFI to force the fuse controller into a state where the OPT_SECURED_BURN_FEATURE_ENABLE (SBFE) bit is cleared, the controller permits arbitrary FF writes.
  • After clearing SBFE, the attacker can write FF rows that set all *_DISABLE option words to zero (or to the defect‑mask), effectively "un‑floorsweeping" the die.
  • The result is a fully unlocked GA100 with up to 128 SMs, 80 GB usable HBM on the 8 GB CMP SKU, PCIe Gen4 x16, NVLink, ECC and MIG support.

GA100 fuse architecture in brief

Region Rows Purpose
Regular fuses (0‑221) Primary/redundant security and product‑segmentation bits. Blowing is permanent and cannot be undone.
Fuseless fuses (FF, 222‑487) Stores JTAG‑style records that map to option‑words. New records replace old ones, allowing software‑style reprogramming.
IFF (487‑512) On‑the‑fly reprogramming; not accessible on production cards.
RIR (extra 8 rows) Manufacturer‑defined repair instructions; largely unusable for unlock on the 8 GB SKU.

The most valuable FF option‑words for unlocking are listed below (chain, block, bit ranges are inferred from reverse‑engineering):

Option word FF location Effect
NV_FUSE_OPT_FBIO_DISABLE Chain 1, bits 83‑106 Disables framebuffer I/O partitions
NV_FUSE_OPT_FBPA_DISABLE Chain 1, bits 107‑130 Disables sub‑partitions within FBPs
NV_FUSE_OPT_GPC_DISABLE Chain 1, bits 131‑138 Disables whole GPCs
NV_FUSE_OPT_TPC_GPC{i}_DISABLE Chain 1, bits 187+8i‑194+8i Disables individual TPCs
NV_FUSE_OPT_NVLINK_DISABLE Chain 1, bits 368‑370 Turns NVLink off
NV_FUSE_OPT_DISABLE_GEN3_SPEED Chain 2, block 39 Limits PCIe to Gen3
NV_FUSE_OPT_PCIE_MAGIC_BITS_D[25] Chain 2, block 47, bit 3 Disables PCIe Gen4
NV_FUSE_OPT_SYS_PIPE_DISABLE Chain 1, bits 372‑379 Blocks MIG pipes

Clearing these bits removes the floorsweep signals that power‑gate the corresponding blocks.


How voltage fault injection enables the unlock

  1. Glitch the PWMVID rail – an STM32‑G4 board monitors NVVDD and injects a short low‑voltage pulse into the VRM PWM control line (PWMVID) during power‑on.
  2. Trigger a “Goldilocks” state – after a few carefully timed pulses, the fuse controller boots with:
    • OPT_SECURED_BURN_FEATURE_ENABLE = 0
    • DISABLE_SW_OVERRIDE = 0
    • ENABLE_FUSE_PROGRAM still at its fused value (1 after a SENSE_CTRL command).
  3. Shadow SBFE – with DISABLE_SW_OVERRIDE = 0, the host can write to the OPT_SECURED_BURN_FEATURE_ENABLE register, permanently setting SBFE to 0 and opening the fuse‑programming path.
  4. Burn FF rows – the attacker writes new FF records that set every *_DISABLE word to the minimum viable mask (defect‑only). The fuse controller accepts these writes because SBFE is cleared.
  5. Finalize configuration – after the FF burn, the standard cmpunlocker driver patches remaining configuration registers (FBPA_CFG, SM_SPEED_SELECT, etc.) to expose the full framebuffer geometry and SM count.

The entire glitch campaign takes between 1 hour and 48 hours depending on hardware stability; each attempt consists of a power‑cycle, a timer‑driven PWM pulse, and a host‑side check of fuse‑controller registers.


Measured performance gains

Metric Retail A100 Software‑only CMP 170HX unlock Hardware‑unlocked CMP 170HX (VFI)
FP16 half2 FMA throughput 72.488 TFLOP/s (1.00×) 50.129 TFLOP/s (0.69×) 85.861 TFLOP/s (1.18×)
SM count 108 (yield‑limited) 70 120‑128 (die‑lottery dependent)
PCIe generation Gen4 x16 (datacenter SKU) Gen2 (limited by FF) Gen4 x16
NVLink Enabled Disabled Enabled after a tiny PCB bridge
Usable HBM 80 GB (A100‑80G) 64 GB (software only) 80 GB

The hardware‑unlocked card consistently outperforms a retail A100 in FP16 workloads, confirming that the extra SMs and higher memory bandwidth are functional.


Practical unlock steps (high‑level)

  1. Prepare the glitch board – flash the STM32 firmware (available in the ga100vfi repo), solder wires to the PWMVID and NVVDD test points on the CMP PCB, and remove the 0402 0 Ω resistor that connects PWMVID to the VRM.
  2. Run the glitch campaign – power‑cycle the card, let the STM32 detect the NVVDD threshold (≈225 mV), and deliver the three‑pulse waveform. Repeat until the controller reports SBFE = 0 and DISABLE_SW_OVERRIDE = 0.
  3. Issue SENSE_CTRL – this forces the controller to latch all regular fuse values while preserving the overridden DISABLE_SW_OVERRIDE.
  4. Shadow SBFE – write 0 to the OPT_SECURED_BURN_FEATURE_ENABLE register via the privileged fuse‑controller interface.
  5. Burn FF rows – program the FF records that clear all *_DISABLE bits (tailored per die based on defect masks). The ga100vfi tool automates row calculation.
  6. Apply software patches – run cmpunlocker to adjust FBPA configuration, SM speed selection, and memory geometry.
  7. Optional hardware mods – bridge the missing U11 pins on the PCB to enable NVLink; add a PCIe‑Gen4 decoupling capacitor if desired.

All scripts and detailed waveforms are in the public GitHub repository: https://github.com/kate-land/ga100vfi.


Community reaction (selected HN comments)

"amazing to think there was a time (literally not too long ago) where you would throw 80 GB of HBM onto a chip and turn 90 % of it off" – @htrp

"But what if NVIDIA used those fuses because an A100 had too many fabrication faults so it could only be a cheaper mining card?" – @amelius

"I’m actually trying to unlock the 80G in a CMP 170HX the software way for the last 2 weeks… seems a lot more risky" – @notenchange

"It's utterly crazy that anyone outside NVIDIA knows any of this…" – @mitxela

These comments highlight both the surprise at the economic rationale behind the segmentation and the perceived risk of hardware‑level fuse blowing.


Future research directions

  • Security‑sensitive fuses – Some option‑words (e.g., debug‑HS execution, secure boot) remain blown in production; targeting them could expose deeper debug interfaces.
  • Glitch‑state automation – Faster glitch campaigns using external PCIe risers or dedicated power‑sequencing rigs could reduce unlock time from hours to minutes.
  • Cross‑SKU applicability – Extending the technique to other GA100‑based GPUs (e.g., Tesla P100, older A100 variants) may reveal similar unlock opportunities.

References

  • Jon Pry, A Canary in the Crypto Mine (2026).
  • Amogh Munikote, cmpunlocker (2026).
  • Otto Bittner et al., The Forgotten Threat of Voltage Glitching (2021).
  • NVIDIA, US9230678B2 – Enhanced fuseless fuse structure (2016).
  • TSMC, TEF07F32X32HD18 fuse datasheet (2018).
  • Kate Land, Using voltage glitching to unlock the A100 (2026) – full blog and source code.

Sources

Related