Skip to content
Featured on Startup Fame

Researcher unlocks 80GB HBM on Nvidia’s CMP 170HX mining card using voltage fault injection

Researcher Kate demonstrated a voltage fault injection attack on Nvidia’s GA100 die to unlock hidden memory and compute resources on the CMP 170HX mining card. The method exposed 80GB of HBM and up to 126 streaming multiprocessors, surpassing the A100’s 108 enabled SMs in a synthetic FP16 benchmark. Nvidia has not responded to the reported silicon-level vulnerability.

Editor, Lazyfounder

Published 4 min read
Researcher unlocks 80GB HBM on Nvidia’s CMP 170HX mining card using voltage fault injection
Image: Nvidia) via source

KEY HIGHLIGHTS

  • Researcher Kate used voltage fault injection to unlock 80GB HBM and up to 126 streaming multiprocessors (SMs) on Nvidia’s CMP 170HX mining card.
  • The CMP 170HX, introduced in 2021 for cryptomining, physically contains the same GA100 die as the A100 but was limited by firmware and fuses.
  • The unlocked card reached 85.9 TFLOP/s in FP16 benchmarks, compared to the A100’s 72.5 TFLOP/s—but tensor performance, ECC, NVLink, and MIG partitioning remain untested.
  • Unlocking a single card takes 1 to 48 hours, with varying success rates and permanent fuse burns.
  • Street prices for the CMP 170HX surged from $250 to over $1,000 following the unlocking discovery.
  • Nvidia has not publicly commented on the vulnerability, which affects already-shipped silicon.

How the unlock works

Kate, an independent researcher, used voltage fault injection to disrupt Nvidia’s GA100 die during boot, allowing her to rewrite fuses designed to be permanent. This technique exposed hidden memory and compute resources on the CMP 170HX, a mining card Nvidia released in 2021 with deliberate limitations to steer harvested dies away from gaming and data center markets.

The CMP 170HX originally shipped with only 8GB or 10GB of HBM2e accessible to software, along with 70 streaming multiprocessors (SMs) and a PCIe Gen1 link. Kate’s method unlocked the full 80GB of HBM and between 120 and 126 SMs, exceeding the A100’s 108 enabled SMs. The attack builds on earlier community efforts like cmpunlocker, which exposed 64GB of memory but left features like NVLink and ECC untouched.

Performance claims and limitations

In a single FP16 fused-multiply-add benchmark, the unlocked CMP 170HX achieved 85.9 TFLOP/s, compared to the A100’s 72.5 TFLOP/s. The gain stems from the higher number of enabled SMs—up to 126 versus the A100’s 108—but the result is narrowly focused.

Critical data center features remain untested or incomplete. Tensor performance, ECC memory protection, NVLink interconnect, and Multi-Instance GPU (MIG) partitioning—all core to the A100’s value—were not validated. NVLink, for example, requires additional hardware modifications to function, and PCIe speeds only reach Gen4 with extra rework.

TechRadar notes that while the FP16 benchmark suggests a performance edge, the lack of testing for other features means the unlocked card cannot yet replace an A100 in production environments. The research instead demonstrates the GA100 die’s native capabilities when freed from Nvidia’s artificial constraints.

Security flaws and market impact

Kate’s research exposes a fundamental security weakness in Nvidia’s Ampere-era silicon: the fuse scheme allows operators to relax restrictions by blowing additional fuses, contrary to security best practices. She disclosed the vulnerability to Nvidia’s product security team before publishing, but the company has not publicly acknowledged the issue. Since the flaw is embedded in silicon, it cannot be patched for cards already in circulation.

The discovery has driven up demand for the CMP 170HX, which sold for as little as $250 before the first software unlock. Prices now exceed $1,000 as buyers seek to exploit the card’s hidden potential. However, unlocking is not guaranteed: out of ten cards Kate examined, one had only 64GB of healthy memory instead of the full 80GB. The process also involves physical rework, per-card tuning, and permanent fuse burns, adding risk and labor for users.

Broader context in hardware security

This research adds to a growing body of work highlighting security vulnerabilities in Nvidia’s hardware. Earlier this year, academic teams demonstrated Rowhammer bit flips on Nvidia GPUs and defeated ECC protections. In June 2026, researchers at Arizona State University, including Jon Pry, published a paper bypassing stack protection in the GPU’s security coprocessor, further exposing risks in Nvidia’s design choices.

What this means

Lazyfounder analysis — our interpretation, not reported fact.

The unlocking of the CMP 170HX reveals how Nvidia’s segmentation strategy—limiting hardware to protect higher-margin products—can be bypassed through physical attacks. For founders and operators in AI and data centers, the story underscores the risks of relying on artificial hardware limitations for protection. While the unlocked card is not a drop-in replacement for an A100, the research demonstrates that last-generation hardware may hold untapped potential, particularly in memory-constrained markets like AI training. The surge in CMP 170HX prices also signals how quickly secondary markets can react to unlocking discoveries, which could influence procurement strategies for cost-sensitive projects.

Key takeaways

  • Nvidia’s artificial limitations on the CMP 170HX can be bypassed using voltage fault injection, exposing hidden memory and compute resources.
  • The unlocked performance advantage is narrow and unproven for critical data center features, making the card unsuitable as an A100 replacement.
  • The vulnerability is rooted in silicon and cannot be fixed for existing cards, posing a long-term security risk for Ampere-era GPUs.
  • Secondary market prices for the CMP 170HX have quadrupled, reflecting demand for unlockable hardware in AI and cryptomining.
  • The research aligns with broader concerns about hardware security, particularly in GPUs used for AI accelerators.

Sources

  1. TechRadar · 2026-11-18
    Voltage glitch unlocks 80GB on Nvidia's CMP 170HX, but 'faster than an A100' rests on a single benchmark

This story is an original summary drafted with AI by Lazyfounder from the reporting listed above and checked by automated validation. Facts are attributed to their original publishers; sections marked as analysis are Lazyfounder's. Where a source is in another language, facts were machine-translated and quotations are reported, not reproduced. Read the original coverage via the links, and see our AI policy and corrections policy.

About the author

Editor, Lazyfounder

Tarun Mottlia edits LazyFounders, covering Indian startups, funding rounds, AI and product launches. Every story on the site is AI-assisted and checked against its cited sources before publication.

More stories by Tarun Mottlia

Get the LazyFounder Brief

Startup, funding and AI news in a five-minute read. Join the early-access list.

Lazy Founder - Powered by Blogy.in

Contact us

Have a story tip, correction or partnership idea?

Write to us at tarun.kumar@blogy.in or talk to the founder directly. We read every message.

Contact us