Richland Technology
Richland Technology
← Back to Blog
open-sourceresearchsecurityhardwaregpu

Voltage-glitching a CMP 170HX into an A100-80G: GA100 fuses that relax instead of restrict, 85.9 vs 72.5 TFLOPS FP16 — unverified, and the HBM map is contested

Kate's voltage fault-injection writeup defeats NVIDIA's one-way eFuse model with an ~$8 STM32 glitcher: 80 GB HBM, 120–128 SMs, NVLink, Gen4 — self-reported, no replication yet

Richland AI Desk·12 November 2026·8 min read
Voltage-glitching a CMP 170HX into an A100-80G: GA100 fuses that relax instead of restrict, 85.9 vs 72.5 TFLOPS FP16 — unverified, and the HBM map is contested

On Nov 10, hardware-security researcher Kate published a writeup showing that voltage fault injection (VFI) against the GA100 die's eFuse controller lets an operator blow arbitrary fuses — and that on the GA100, blowing fuses relaxes restrictions instead of tightening them. She used it to turn an 8 GB CMP 170HX Ethereum mining card into what she calls "effectively an A100-80G": the full 80 GB of HBM usable, 120–126 SMs enabled, and 85.858 TFLOP/s FP16 measured via OpenCL against 72.488 on a retail A100 — about 1.18× faster (blog.kate.land).

What shipped

The victim card. A CMP 170HX is a GA100-based mining card: 70 SMs (4,480 CUDA cores), 8 GB of application-visible HBM2e (a 10 GB sibling exists), passive cooling, and PCIe that ships firmware-locked to Gen1 — launched September 2021 at a $4,299 MSRP, per the community-maintained 170th Street spec page. The GA100 die behind it is the A100's: 128 SMs as fabbed, a PCIe Gen4 x16 controller, and NVLink links (170th Street, Wikipedia on the A100). Kate's writeup claims the 8 GB card physically carries ~80 GB of HBM and that nearly all of the crippling is eFuse-driven — fuses that disable memory I/O, framebuffer subpartitions, GPCs, NVLink, and the rest.

The mechanism is an inverted fuse model. NVIDIA's segmentation assumes one-way fuses. As Kate puts it: "In a properly secured design, a blown fuse should always induce a more restrictive state, because a blown fuse cannot be unblown in an economically feasible way. NVIDIA's design allows one to relax restrictions by blowing more fuses!"

The GA100 fuse controller holds 16,384 eFuses in 512 rows of 32. In the "fuseless" (FF) rows 222–487, option words aren't individual fuse bits — they're a JTAG-chain-like script, and when two records address the same chain/block, the later record wins. Appending a row replaces the effective value of a 16-bit block without physically unblowing anything. That makes FF-backed options reversible by writing more fuses — the opposite of a one-way gate. NVIDIA knows this, which is why, "under normal conditions, you are strictly prevented from performing any FF fuse blows."

The glitcher. VFI is how Kate gets past that write-protection. Her rig: an STM32G4 board (STM32G431 per the repo) that taps the PWMVID signal on the card's PCB and drives the MP2888B voltage regulator with timed waveforms, after removing a single 0402 resistor so the regulator follows the glitcher instead of the GPU (LavX News, Hardware Busters). Glitches fire during boot, timed off NVVDD crossing 225 mV, and the host machine cycles between ACPI S0 and S3 states to replay each attempt. Tom's Hardware reports the glitcher PCB around the ~$8 microcontroller was itself designed with the help of an AI model (Tom's Hardware). Per-card economics: one to 48 hours of glitching, across ten cards — nine came out with the full 80 GB, one with only 64 GB intact.

What unlocks. Once a glitch lands, the fuse controller's own protection words read as unblown — DISABLE_SW_OVERRIDE at zero, ENABLE_FUSE_PROGRAM at zero — so a SENSE_CTRL command restores most control-fuse options while leaving software override available (LavX's summary of the fault state). From there Kate shadowed OPT_SECURED_BURN_FEATURE_ENABLE to zero and wrote new FF rows whose disable masks match what each die needs disabled. Her results table: 64–80 GB framebuffer, 120–128 SMs, PCIe Gen4, NVLink and ECC — with MIG, NVJPG, and NVDEC marked unlockable but untested. Her two unlocked cards appear in nvidia-smi as "NVIDIA Graphics Device" with 81,920 MiB each, device ID 0x20C210DE, connected via NV12 (NVLink) topology.

Measured performance. The OpenCL run (the ProjectPhysX benchmark, per Hardware Busters) reports 85.858 TFLOP/s FP16 FMA at 124 compute units / 1,410 MHz, versus 72.488 for her retail A100 in the same kernel (1.18×) and 50.129 for a software-unlocked CMP 170HX (per LavX) — about 1.7× over the cmpunlocker state, per Hardware Busters. Coalesced memory bandwidth reads ~1,883 GB/s and writes ~2,138 GB/s — in A100 territory for an 80 GB HBM part. The unlocked card outruns the flagship mostly because retail A100s are locked to the yield-safe 108 SMs (6,912 CUDA cores), while her unlock runs every functional SM — 120–126, "die-lottery dependent." Per Hardware Busters, she says real inference stacks such as vLLM run across the full 80 GB.

Disclosure and code. Kate says the findings were disclosed to NVIDIA's product security team ahead of publication. Her work builds on Jon Pry's earlier GA100 research and on Amogh Munikote's cmpunlocker software unlocks, which reached 64 GB / PCIe Gen2; another researcher — unnamed in her post — independently found the same vulnerability. She published the infrastructure as github.com/kate-land/ga100vfi: the PWMVID control-board design, STM32 firmware, host tools, GA100 register-capture code, and the experiment-loop description, MIT-licensed.

Why it matters

The durable lesson is that NVIDIA's product segmentation is only as strong as the fuse controller's write-protection. On the GA100, one die becomes everything from a mining orphan to an A100-class datacenter part purely by option words — and the security assumption that fuses only ever restrict is what allows those option words to be reversible by append. Defeat the controller's gate once, with a ~$8 microcontroller and patience, and the segmentation scheme for an entire die family ("almost 20 different SKUs," per Kate's post) collapses in one go.

That flips the used-market calculus for local-inference builders who buy crippled ex-mining cards. The 9-of-10 fully working result is the strongest evidence yet that the memory crippling is market segmentation rather than defect sorting — though HN commenters dispute this, with some reporting unlocked 80 GB cards that carry bad HBM (HN thread). If it replicates, the used market gains an 80 GB high-bandwidth GPU at mining-card prices: LTT Labs' September comparison table put used CMP 170HX cards in the $1,300–1,700 range (LTT Labs), and HN commenters say prices have already jumped.

For the audience building on GPUs, the practical read is narrower: this is not a weekend project, and the payoff is a card whose measured FP16 beats a retail A100's by 18%, with NVLink available for multi-card clusters. The security takeaway, though, is general: if you ship silicon whose value tiers rest on eFuses, the fuse controller is now a legitimate attack surface, and VFI is the tool.

Caveats

Everything above that is not community corroboration is Kate's own report — published two days ago, by a single operator on ten cards, with no independent replication yet. Treat the headline numbers as self-reported, not independently confirmed.

  • The A100 comparison is a same-kernel measurement, not an NVIDIA spec-sheet number. The 72.488 TFLOP/s figure comes from running the same ProjectPhysX OpenCL kernel on a retail A100, not from NVIDIA's published tensor throughput.
  • The biggest factual conflict is the hardware itself. Kate says the 8 GB SKU carries ~80 GB of HBM on board. Community resources built on VBIOS dumps say the opposite: LTT Labs, citing TechPowerUp, lists the 8 GB card as carrying 64 GB of physical HBM2e and the 10 GB card 80 GB — the reverse mapping — while 170th Street's page lists the unlocked capacity as 64 GB/40 GB rather than Kate's 80 GB. If they're right, Kate's "partially defective" 64 GB card may simply be an 8 GB card with fewer working stacks, and her 9/10 result still needs independent confirmation.
  • Internal inconsistency in her own output. Her OpenCL log reports PCIe Gen3 x16 bandwidth (11.58 GB/s bidirectional) while she headlines Gen4 as restored.
  • ECC is claimed, not demonstrated. Her table marks ECC restored, but no ECC test output appears in the published writeup; 170th Street lists CMP ECC as disabled and LTT Labs found on-die ECC "yet to be exposed" by the software unlock.
  • Fuses are irreversible, and failed glitch states are destructive. Some glitch states left GPUs "in an unusable condition" (LavX), defective SMs must be correctly identified and masked or the die is permanently lost, and on the 8 GB SKU 15 of 16 RIR repair words are all-ones, leaving almost no fuse-level repair path.
  • Physical-modding risks. The mod removes an onboard resistor, and Kate warns that enabling NVLink ties together the 3.3 V rails of cards that may sit on different power supplies (Hardware Busters). The card is passively cooled; LTT Labs found that at unlocked compute it needed aggressive aftermarket fans to stay under the 85 °C limit (measured in the software-unlock state; the VFI writeup publishes no unlocked power-draw figures).
  • Software is still in the loop. Tom's notes the CMPUnlocker driver mod is still required for the newly enabled features to work in the driver; the blown fuses themselves are physical and persistent.

Sources

Start a Conversation

Ready to build something like this?

Tell us about your project and we'll respond with a clear plan within one business day.

Tell Us About Your Project

Takes 3 minutes. We read every submission personally.